1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
20 // Terminology note: this code has a lot of handling for "post-increment" or
21 // "post-inc" users. This is not talking about post-increment addressing modes;
22 // it is instead talking about code like this:
24 // %i = phi [ 0, %entry ], [ %i.next, %latch ]
26 // %i.next = add %i, 1
27 // %c = icmp eq %i.next, %n
29 // The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however
30 // it's useful to think about these as the same register, with some uses using
31 // the value of the register before the add and some using // it after. In this
32 // example, the icmp is a post-increment user, since it uses %i.next, which is
33 // the value of the induction variable after the increment. The other common
34 // case of post-increment users is users outside the loop.
36 // TODO: More sophistication in the way Formulae are generated and filtered.
38 // TODO: Handle multiple loops at a time.
40 // TODO: Should the addressing mode BaseGV be changed to a ConstantExpr instead
43 // TODO: When truncation is free, truncate ICmp users' operands to make it a
44 // smaller encoding (on x86 at least).
46 // TODO: When a negated register is used by an add (such as in a list of
47 // multiple base registers, or as the increment expression in an addrec),
48 // we may not actually need both reg and (-1 * reg) in registers; the
49 // negation can be implemented by using a sub instead of an add. The
50 // lack of support for taking this into consideration when making
51 // register pressure decisions is partly worked around by the "Special"
54 //===----------------------------------------------------------------------===//
56 #define DEBUG_TYPE "loop-reduce"
57 #include "llvm/Transforms/Scalar.h"
58 #include "llvm/ADT/DenseSet.h"
59 #include "llvm/ADT/SetVector.h"
60 #include "llvm/ADT/SmallBitVector.h"
61 #include "llvm/ADT/STLExtras.h"
62 #include "llvm/Analysis/Dominators.h"
63 #include "llvm/Analysis/IVUsers.h"
64 #include "llvm/Analysis/LoopPass.h"
65 #include "llvm/Analysis/ScalarEvolutionExpander.h"
66 #include "llvm/Analysis/TargetTransformInfo.h"
67 #include "llvm/Assembly/Writer.h"
68 #include "llvm/IR/Constants.h"
69 #include "llvm/IR/DerivedTypes.h"
70 #include "llvm/IR/Instructions.h"
71 #include "llvm/IR/IntrinsicInst.h"
72 #include "llvm/Support/CommandLine.h"
73 #include "llvm/Support/Debug.h"
74 #include "llvm/Support/ValueHandle.h"
75 #include "llvm/Support/raw_ostream.h"
76 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
77 #include "llvm/Transforms/Utils/Local.h"
81 /// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for
82 /// bail out. This threshold is far beyond the number of users that LSR can
83 /// conceivably solve, so it should not affect generated code, but catches the
84 /// worst cases before LSR burns too much compile time and stack space.
85 static const unsigned MaxIVUsers = 200;
87 // Temporary flag to cleanup congruent phis after LSR phi expansion.
88 // It's currently disabled until we can determine whether it's truly useful or
89 // not. The flag should be removed after the v3.0 release.
90 // This is now needed for ivchains.
91 static cl::opt<bool> EnablePhiElim(
92 "enable-lsr-phielim", cl::Hidden, cl::init(true),
93 cl::desc("Enable LSR phi elimination"));
96 // Stress test IV chain generation.
97 static cl::opt<bool> StressIVChain(
98 "stress-ivchain", cl::Hidden, cl::init(false),
99 cl::desc("Stress test LSR IV chains"));
101 static bool StressIVChain = false;
106 /// RegSortData - This class holds data which is used to order reuse candidates.
109 /// UsedByIndices - This represents the set of LSRUse indices which reference
110 /// a particular register.
111 SmallBitVector UsedByIndices;
115 void print(raw_ostream &OS) const;
121 void RegSortData::print(raw_ostream &OS) const {
122 OS << "[NumUses=" << UsedByIndices.count() << ']';
125 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
126 void RegSortData::dump() const {
127 print(errs()); errs() << '\n';
133 /// RegUseTracker - Map register candidates to information about how they are
135 class RegUseTracker {
136 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
138 RegUsesTy RegUsesMap;
139 SmallVector<const SCEV *, 16> RegSequence;
142 void CountRegister(const SCEV *Reg, size_t LUIdx);
143 void DropRegister(const SCEV *Reg, size_t LUIdx);
144 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
146 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
148 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
152 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
153 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
154 iterator begin() { return RegSequence.begin(); }
155 iterator end() { return RegSequence.end(); }
156 const_iterator begin() const { return RegSequence.begin(); }
157 const_iterator end() const { return RegSequence.end(); }
163 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
164 std::pair<RegUsesTy::iterator, bool> Pair =
165 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
166 RegSortData &RSD = Pair.first->second;
168 RegSequence.push_back(Reg);
169 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
170 RSD.UsedByIndices.set(LUIdx);
174 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
175 RegUsesTy::iterator It = RegUsesMap.find(Reg);
176 assert(It != RegUsesMap.end());
177 RegSortData &RSD = It->second;
178 assert(RSD.UsedByIndices.size() > LUIdx);
179 RSD.UsedByIndices.reset(LUIdx);
183 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
184 assert(LUIdx <= LastLUIdx);
186 // Update RegUses. The data structure is not optimized for this purpose;
187 // we must iterate through it and update each of the bit vectors.
188 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
190 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
191 if (LUIdx < UsedByIndices.size())
192 UsedByIndices[LUIdx] =
193 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
194 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
199 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
200 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
201 if (I == RegUsesMap.end())
203 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
204 int i = UsedByIndices.find_first();
205 if (i == -1) return false;
206 if ((size_t)i != LUIdx) return true;
207 return UsedByIndices.find_next(i) != -1;
210 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
211 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
212 assert(I != RegUsesMap.end() && "Unknown register!");
213 return I->second.UsedByIndices;
216 void RegUseTracker::clear() {
223 /// Formula - This class holds information that describes a formula for
224 /// computing satisfying a use. It may include broken-out immediates and scaled
227 /// Global base address used for complex addressing.
230 /// Base offset for complex addressing.
233 /// Whether any complex addressing has a base register.
236 /// The scale of any complex addressing.
239 /// BaseRegs - The list of "base" registers for this use. When this is
241 SmallVector<const SCEV *, 4> BaseRegs;
243 /// ScaledReg - The 'scaled' register for this use. This should be non-null
244 /// when Scale is not zero.
245 const SCEV *ScaledReg;
247 /// UnfoldedOffset - An additional constant offset which added near the
248 /// use. This requires a temporary register, but the offset itself can
249 /// live in an add immediate field rather than a register.
250 int64_t UnfoldedOffset;
253 : BaseGV(0), BaseOffset(0), HasBaseReg(false), Scale(0), ScaledReg(0),
256 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
258 unsigned getNumRegs() const;
259 Type *getType() const;
261 void DeleteBaseReg(const SCEV *&S);
263 bool referencesReg(const SCEV *S) const;
264 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
265 const RegUseTracker &RegUses) const;
267 void print(raw_ostream &OS) const;
273 /// DoInitialMatch - Recursion helper for InitialMatch.
274 static void DoInitialMatch(const SCEV *S, Loop *L,
275 SmallVectorImpl<const SCEV *> &Good,
276 SmallVectorImpl<const SCEV *> &Bad,
277 ScalarEvolution &SE) {
278 // Collect expressions which properly dominate the loop header.
279 if (SE.properlyDominates(S, L->getHeader())) {
284 // Look at add operands.
285 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
286 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
288 DoInitialMatch(*I, L, Good, Bad, SE);
292 // Look at addrec operands.
293 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
294 if (!AR->getStart()->isZero()) {
295 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
296 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
297 AR->getStepRecurrence(SE),
298 // FIXME: AR->getNoWrapFlags()
299 AR->getLoop(), SCEV::FlagAnyWrap),
304 // Handle a multiplication by -1 (negation) if it didn't fold.
305 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
306 if (Mul->getOperand(0)->isAllOnesValue()) {
307 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
308 const SCEV *NewMul = SE.getMulExpr(Ops);
310 SmallVector<const SCEV *, 4> MyGood;
311 SmallVector<const SCEV *, 4> MyBad;
312 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
313 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
314 SE.getEffectiveSCEVType(NewMul->getType())));
315 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
316 E = MyGood.end(); I != E; ++I)
317 Good.push_back(SE.getMulExpr(NegOne, *I));
318 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
319 E = MyBad.end(); I != E; ++I)
320 Bad.push_back(SE.getMulExpr(NegOne, *I));
324 // Ok, we can't do anything interesting. Just stuff the whole thing into a
325 // register and hope for the best.
329 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
330 /// attempting to keep all loop-invariant and loop-computable values in a
331 /// single base register.
332 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
333 SmallVector<const SCEV *, 4> Good;
334 SmallVector<const SCEV *, 4> Bad;
335 DoInitialMatch(S, L, Good, Bad, SE);
337 const SCEV *Sum = SE.getAddExpr(Good);
339 BaseRegs.push_back(Sum);
343 const SCEV *Sum = SE.getAddExpr(Bad);
345 BaseRegs.push_back(Sum);
350 /// getNumRegs - Return the total number of register operands used by this
351 /// formula. This does not include register uses implied by non-constant
353 unsigned Formula::getNumRegs() const {
354 return !!ScaledReg + BaseRegs.size();
357 /// getType - Return the type of this formula, if it has one, or null
358 /// otherwise. This type is meaningless except for the bit size.
359 Type *Formula::getType() const {
360 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
361 ScaledReg ? ScaledReg->getType() :
362 BaseGV ? BaseGV->getType() :
366 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
367 void Formula::DeleteBaseReg(const SCEV *&S) {
368 if (&S != &BaseRegs.back())
369 std::swap(S, BaseRegs.back());
373 /// referencesReg - Test if this formula references the given register.
374 bool Formula::referencesReg(const SCEV *S) const {
375 return S == ScaledReg ||
376 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
379 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
380 /// which are used by uses other than the use with the given index.
381 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
382 const RegUseTracker &RegUses) const {
384 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
386 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
387 E = BaseRegs.end(); I != E; ++I)
388 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
393 void Formula::print(raw_ostream &OS) const {
396 if (!First) OS << " + "; else First = false;
397 WriteAsOperand(OS, BaseGV, /*PrintType=*/false);
399 if (BaseOffset != 0) {
400 if (!First) OS << " + "; else First = false;
403 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
404 E = BaseRegs.end(); I != E; ++I) {
405 if (!First) OS << " + "; else First = false;
406 OS << "reg(" << **I << ')';
408 if (HasBaseReg && BaseRegs.empty()) {
409 if (!First) OS << " + "; else First = false;
410 OS << "**error: HasBaseReg**";
411 } else if (!HasBaseReg && !BaseRegs.empty()) {
412 if (!First) OS << " + "; else First = false;
413 OS << "**error: !HasBaseReg**";
416 if (!First) OS << " + "; else First = false;
417 OS << Scale << "*reg(";
424 if (UnfoldedOffset != 0) {
425 if (!First) OS << " + "; else First = false;
426 OS << "imm(" << UnfoldedOffset << ')';
430 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
431 void Formula::dump() const {
432 print(errs()); errs() << '\n';
436 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
437 /// without changing its value.
438 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
440 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
441 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
444 /// isAddSExtable - Return true if the given add can be sign-extended
445 /// without changing its value.
446 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
448 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
449 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
452 /// isMulSExtable - Return true if the given mul can be sign-extended
453 /// without changing its value.
454 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
456 IntegerType::get(SE.getContext(),
457 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
458 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
461 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
462 /// and if the remainder is known to be zero, or null otherwise. If
463 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
464 /// to Y, ignoring that the multiplication may overflow, which is useful when
465 /// the result will be used in a context where the most significant bits are
467 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
469 bool IgnoreSignificantBits = false) {
470 // Handle the trivial case, which works for any SCEV type.
472 return SE.getConstant(LHS->getType(), 1);
474 // Handle a few RHS special cases.
475 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
477 const APInt &RA = RC->getValue()->getValue();
478 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
480 if (RA.isAllOnesValue())
481 return SE.getMulExpr(LHS, RC);
482 // Handle x /s 1 as x.
487 // Check for a division of a constant by a constant.
488 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
491 const APInt &LA = C->getValue()->getValue();
492 const APInt &RA = RC->getValue()->getValue();
493 if (LA.srem(RA) != 0)
495 return SE.getConstant(LA.sdiv(RA));
498 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
499 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
500 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
501 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
502 IgnoreSignificantBits);
504 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
505 IgnoreSignificantBits);
506 if (!Start) return 0;
507 // FlagNW is independent of the start value, step direction, and is
508 // preserved with smaller magnitude steps.
509 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
510 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
515 // Distribute the sdiv over add operands, if the add doesn't overflow.
516 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
517 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
518 SmallVector<const SCEV *, 8> Ops;
519 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
521 const SCEV *Op = getExactSDiv(*I, RHS, SE,
522 IgnoreSignificantBits);
526 return SE.getAddExpr(Ops);
531 // Check for a multiply operand that we can pull RHS out of.
532 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
533 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
534 SmallVector<const SCEV *, 4> Ops;
536 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
540 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
541 IgnoreSignificantBits)) {
547 return Found ? SE.getMulExpr(Ops) : 0;
552 // Otherwise we don't know.
556 /// ExtractImmediate - If S involves the addition of a constant integer value,
557 /// return that integer value, and mutate S to point to a new SCEV with that
559 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
560 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
561 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
562 S = SE.getConstant(C->getType(), 0);
563 return C->getValue()->getSExtValue();
565 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
566 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
567 int64_t Result = ExtractImmediate(NewOps.front(), SE);
569 S = SE.getAddExpr(NewOps);
571 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
572 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
573 int64_t Result = ExtractImmediate(NewOps.front(), SE);
575 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
576 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
583 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
584 /// return that symbol, and mutate S to point to a new SCEV with that
586 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
587 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
588 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
589 S = SE.getConstant(GV->getType(), 0);
592 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
593 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
594 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
596 S = SE.getAddExpr(NewOps);
598 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
599 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
600 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
602 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
603 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
610 /// isAddressUse - Returns true if the specified instruction is using the
611 /// specified value as an address.
612 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
613 bool isAddress = isa<LoadInst>(Inst);
614 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
615 if (SI->getOperand(1) == OperandVal)
617 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
618 // Addressing modes can also be folded into prefetches and a variety
620 switch (II->getIntrinsicID()) {
622 case Intrinsic::prefetch:
623 case Intrinsic::x86_sse_storeu_ps:
624 case Intrinsic::x86_sse2_storeu_pd:
625 case Intrinsic::x86_sse2_storeu_dq:
626 case Intrinsic::x86_sse2_storel_dq:
627 if (II->getArgOperand(0) == OperandVal)
635 /// getAccessType - Return the type of the memory being accessed.
636 static Type *getAccessType(const Instruction *Inst) {
637 Type *AccessTy = Inst->getType();
638 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
639 AccessTy = SI->getOperand(0)->getType();
640 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
641 // Addressing modes can also be folded into prefetches and a variety
643 switch (II->getIntrinsicID()) {
645 case Intrinsic::x86_sse_storeu_ps:
646 case Intrinsic::x86_sse2_storeu_pd:
647 case Intrinsic::x86_sse2_storeu_dq:
648 case Intrinsic::x86_sse2_storel_dq:
649 AccessTy = II->getArgOperand(0)->getType();
654 // All pointers have the same requirements, so canonicalize them to an
655 // arbitrary pointer type to minimize variation.
656 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy))
657 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
658 PTy->getAddressSpace());
663 /// isExistingPhi - Return true if this AddRec is already a phi in its loop.
664 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
665 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
666 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
667 if (SE.isSCEVable(PN->getType()) &&
668 (SE.getEffectiveSCEVType(PN->getType()) ==
669 SE.getEffectiveSCEVType(AR->getType())) &&
670 SE.getSCEV(PN) == AR)
676 /// Check if expanding this expression is likely to incur significant cost. This
677 /// is tricky because SCEV doesn't track which expressions are actually computed
678 /// by the current IR.
680 /// We currently allow expansion of IV increments that involve adds,
681 /// multiplication by constants, and AddRecs from existing phis.
683 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
684 /// obvious multiple of the UDivExpr.
685 static bool isHighCostExpansion(const SCEV *S,
686 SmallPtrSet<const SCEV*, 8> &Processed,
687 ScalarEvolution &SE) {
688 // Zero/One operand expressions
689 switch (S->getSCEVType()) {
694 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
697 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
700 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
704 if (!Processed.insert(S))
707 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
708 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
710 if (isHighCostExpansion(*I, Processed, SE))
716 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
717 if (Mul->getNumOperands() == 2) {
718 // Multiplication by a constant is ok
719 if (isa<SCEVConstant>(Mul->getOperand(0)))
720 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
722 // If we have the value of one operand, check if an existing
723 // multiplication already generates this expression.
724 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
725 Value *UVal = U->getValue();
726 for (Value::use_iterator UI = UVal->use_begin(), UE = UVal->use_end();
728 // If U is a constant, it may be used by a ConstantExpr.
729 Instruction *User = dyn_cast<Instruction>(*UI);
730 if (User && User->getOpcode() == Instruction::Mul
731 && SE.isSCEVable(User->getType())) {
732 return SE.getSCEV(User) == Mul;
739 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
740 if (isExistingPhi(AR, SE))
744 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
748 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
749 /// specified set are trivially dead, delete them and see if this makes any of
750 /// their operands subsequently dead.
752 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
753 bool Changed = false;
755 while (!DeadInsts.empty()) {
756 Value *V = DeadInsts.pop_back_val();
757 Instruction *I = dyn_cast_or_null<Instruction>(V);
759 if (I == 0 || !isInstructionTriviallyDead(I))
762 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
763 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
766 DeadInsts.push_back(U);
769 I->eraseFromParent();
778 /// Cost - This class is used to measure and compare candidate formulae.
780 /// TODO: Some of these could be merged. Also, a lexical ordering
781 /// isn't always optimal.
785 unsigned NumBaseAdds;
791 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
794 bool operator<(const Cost &Other) const;
799 // Once any of the metrics loses, they must all remain losers.
801 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
802 | ImmCost | SetupCost) != ~0u)
803 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
804 & ImmCost & SetupCost) == ~0u);
809 assert(isValid() && "invalid cost");
810 return NumRegs == ~0u;
813 void RateFormula(const Formula &F,
814 SmallPtrSet<const SCEV *, 16> &Regs,
815 const DenseSet<const SCEV *> &VisitedRegs,
817 const SmallVectorImpl<int64_t> &Offsets,
818 ScalarEvolution &SE, DominatorTree &DT,
819 SmallPtrSet<const SCEV *, 16> *LoserRegs = 0);
821 void print(raw_ostream &OS) const;
825 void RateRegister(const SCEV *Reg,
826 SmallPtrSet<const SCEV *, 16> &Regs,
828 ScalarEvolution &SE, DominatorTree &DT);
829 void RatePrimaryRegister(const SCEV *Reg,
830 SmallPtrSet<const SCEV *, 16> &Regs,
832 ScalarEvolution &SE, DominatorTree &DT,
833 SmallPtrSet<const SCEV *, 16> *LoserRegs);
838 /// RateRegister - Tally up interesting quantities from the given register.
839 void Cost::RateRegister(const SCEV *Reg,
840 SmallPtrSet<const SCEV *, 16> &Regs,
842 ScalarEvolution &SE, DominatorTree &DT) {
843 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
844 // If this is an addrec for another loop, don't second-guess its addrec phi
845 // nodes. LSR isn't currently smart enough to reason about more than one
846 // loop at a time. LSR has already run on inner loops, will not run on outer
847 // loops, and cannot be expected to change sibling loops.
848 if (AR->getLoop() != L) {
849 // If the AddRec exists, consider it's register free and leave it alone.
850 if (isExistingPhi(AR, SE))
853 // Otherwise, do not consider this formula at all.
857 AddRecCost += 1; /// TODO: This should be a function of the stride.
859 // Add the step value register, if it needs one.
860 // TODO: The non-affine case isn't precisely modeled here.
861 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
862 if (!Regs.count(AR->getOperand(1))) {
863 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
871 // Rough heuristic; favor registers which don't require extra setup
872 // instructions in the preheader.
873 if (!isa<SCEVUnknown>(Reg) &&
874 !isa<SCEVConstant>(Reg) &&
875 !(isa<SCEVAddRecExpr>(Reg) &&
876 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
877 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
880 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
881 SE.hasComputableLoopEvolution(Reg, L);
884 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
885 /// before, rate it. Optional LoserRegs provides a way to declare any formula
886 /// that refers to one of those regs an instant loser.
887 void Cost::RatePrimaryRegister(const SCEV *Reg,
888 SmallPtrSet<const SCEV *, 16> &Regs,
890 ScalarEvolution &SE, DominatorTree &DT,
891 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
892 if (LoserRegs && LoserRegs->count(Reg)) {
896 if (Regs.insert(Reg)) {
897 RateRegister(Reg, Regs, L, SE, DT);
898 if (LoserRegs && isLoser())
899 LoserRegs->insert(Reg);
903 void Cost::RateFormula(const Formula &F,
904 SmallPtrSet<const SCEV *, 16> &Regs,
905 const DenseSet<const SCEV *> &VisitedRegs,
907 const SmallVectorImpl<int64_t> &Offsets,
908 ScalarEvolution &SE, DominatorTree &DT,
909 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
910 // Tally up the registers.
911 if (const SCEV *ScaledReg = F.ScaledReg) {
912 if (VisitedRegs.count(ScaledReg)) {
916 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
920 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
921 E = F.BaseRegs.end(); I != E; ++I) {
922 const SCEV *BaseReg = *I;
923 if (VisitedRegs.count(BaseReg)) {
927 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
932 // Determine how many (unfolded) adds we'll need inside the loop.
933 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
934 if (NumBaseParts > 1)
935 NumBaseAdds += NumBaseParts - 1;
937 // Tally up the non-zero immediates.
938 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
939 E = Offsets.end(); I != E; ++I) {
940 int64_t Offset = (uint64_t)*I + F.BaseOffset;
942 ImmCost += 64; // Handle symbolic values conservatively.
943 // TODO: This should probably be the pointer size.
944 else if (Offset != 0)
945 ImmCost += APInt(64, Offset, true).getMinSignedBits();
947 assert(isValid() && "invalid cost");
950 /// Loose - Set this cost to a losing value.
960 /// operator< - Choose the lower cost.
961 bool Cost::operator<(const Cost &Other) const {
962 if (NumRegs != Other.NumRegs)
963 return NumRegs < Other.NumRegs;
964 if (AddRecCost != Other.AddRecCost)
965 return AddRecCost < Other.AddRecCost;
966 if (NumIVMuls != Other.NumIVMuls)
967 return NumIVMuls < Other.NumIVMuls;
968 if (NumBaseAdds != Other.NumBaseAdds)
969 return NumBaseAdds < Other.NumBaseAdds;
970 if (ImmCost != Other.ImmCost)
971 return ImmCost < Other.ImmCost;
972 if (SetupCost != Other.SetupCost)
973 return SetupCost < Other.SetupCost;
977 void Cost::print(raw_ostream &OS) const {
978 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
980 OS << ", with addrec cost " << AddRecCost;
982 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
983 if (NumBaseAdds != 0)
984 OS << ", plus " << NumBaseAdds << " base add"
985 << (NumBaseAdds == 1 ? "" : "s");
987 OS << ", plus " << ImmCost << " imm cost";
989 OS << ", plus " << SetupCost << " setup cost";
992 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
993 void Cost::dump() const {
994 print(errs()); errs() << '\n';
1000 /// LSRFixup - An operand value in an instruction which is to be replaced
1001 /// with some equivalent, possibly strength-reduced, replacement.
1003 /// UserInst - The instruction which will be updated.
1004 Instruction *UserInst;
1006 /// OperandValToReplace - The operand of the instruction which will
1007 /// be replaced. The operand may be used more than once; every instance
1008 /// will be replaced.
1009 Value *OperandValToReplace;
1011 /// PostIncLoops - If this user is to use the post-incremented value of an
1012 /// induction variable, this variable is non-null and holds the loop
1013 /// associated with the induction variable.
1014 PostIncLoopSet PostIncLoops;
1016 /// LUIdx - The index of the LSRUse describing the expression which
1017 /// this fixup needs, minus an offset (below).
1020 /// Offset - A constant offset to be added to the LSRUse expression.
1021 /// This allows multiple fixups to share the same LSRUse with different
1022 /// offsets, for example in an unrolled loop.
1025 bool isUseFullyOutsideLoop(const Loop *L) const;
1029 void print(raw_ostream &OS) const;
1035 LSRFixup::LSRFixup()
1036 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
1038 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
1039 /// value outside of the given loop.
1040 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1041 // PHI nodes use their value in their incoming blocks.
1042 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1043 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1044 if (PN->getIncomingValue(i) == OperandValToReplace &&
1045 L->contains(PN->getIncomingBlock(i)))
1050 return !L->contains(UserInst);
1053 void LSRFixup::print(raw_ostream &OS) const {
1055 // Store is common and interesting enough to be worth special-casing.
1056 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1058 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
1059 } else if (UserInst->getType()->isVoidTy())
1060 OS << UserInst->getOpcodeName();
1062 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
1064 OS << ", OperandValToReplace=";
1065 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
1067 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
1068 E = PostIncLoops.end(); I != E; ++I) {
1069 OS << ", PostIncLoop=";
1070 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
1073 if (LUIdx != ~size_t(0))
1074 OS << ", LUIdx=" << LUIdx;
1077 OS << ", Offset=" << Offset;
1080 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1081 void LSRFixup::dump() const {
1082 print(errs()); errs() << '\n';
1088 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
1089 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
1090 struct UniquifierDenseMapInfo {
1091 static SmallVector<const SCEV *, 4> getEmptyKey() {
1092 SmallVector<const SCEV *, 4> V;
1093 V.push_back(reinterpret_cast<const SCEV *>(-1));
1097 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1098 SmallVector<const SCEV *, 4> V;
1099 V.push_back(reinterpret_cast<const SCEV *>(-2));
1103 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1104 unsigned Result = 0;
1105 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
1106 E = V.end(); I != E; ++I)
1107 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
1111 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1112 const SmallVector<const SCEV *, 4> &RHS) {
1117 /// LSRUse - This class holds the state that LSR keeps for each use in
1118 /// IVUsers, as well as uses invented by LSR itself. It includes information
1119 /// about what kinds of things can be folded into the user, information about
1120 /// the user itself, and information about how the use may be satisfied.
1121 /// TODO: Represent multiple users of the same expression in common?
1123 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1126 /// KindType - An enum for a kind of use, indicating what types of
1127 /// scaled and immediate operands it might support.
1129 Basic, ///< A normal use, with no folding.
1130 Special, ///< A special case of basic, allowing -1 scales.
1131 Address, ///< An address use; folding according to TargetLowering
1132 ICmpZero ///< An equality icmp with both operands folded into one.
1133 // TODO: Add a generic icmp too?
1139 SmallVector<int64_t, 8> Offsets;
1143 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1144 /// LSRUse are outside of the loop, in which case some special-case heuristics
1146 bool AllFixupsOutsideLoop;
1148 /// WidestFixupType - This records the widest use type for any fixup using
1149 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1150 /// max fixup widths to be equivalent, because the narrower one may be relying
1151 /// on the implicit truncation to truncate away bogus bits.
1152 Type *WidestFixupType;
1154 /// Formulae - A list of ways to build a value that can satisfy this user.
1155 /// After the list is populated, one of these is selected heuristically and
1156 /// used to formulate a replacement for OperandValToReplace in UserInst.
1157 SmallVector<Formula, 12> Formulae;
1159 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1160 SmallPtrSet<const SCEV *, 4> Regs;
1162 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1163 MinOffset(INT64_MAX),
1164 MaxOffset(INT64_MIN),
1165 AllFixupsOutsideLoop(true),
1166 WidestFixupType(0) {}
1168 bool HasFormulaWithSameRegs(const Formula &F) const;
1169 bool InsertFormula(const Formula &F);
1170 void DeleteFormula(Formula &F);
1171 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1173 void print(raw_ostream &OS) const;
1179 /// HasFormula - Test whether this use as a formula which has the same
1180 /// registers as the given formula.
1181 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1182 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1183 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1184 // Unstable sort by host order ok, because this is only used for uniquifying.
1185 std::sort(Key.begin(), Key.end());
1186 return Uniquifier.count(Key);
1189 /// InsertFormula - If the given formula has not yet been inserted, add it to
1190 /// the list, and return true. Return false otherwise.
1191 bool LSRUse::InsertFormula(const Formula &F) {
1192 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1193 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1194 // Unstable sort by host order ok, because this is only used for uniquifying.
1195 std::sort(Key.begin(), Key.end());
1197 if (!Uniquifier.insert(Key).second)
1200 // Using a register to hold the value of 0 is not profitable.
1201 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1202 "Zero allocated in a scaled register!");
1204 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1205 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1206 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1209 // Add the formula to the list.
1210 Formulae.push_back(F);
1212 // Record registers now being used by this use.
1213 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1218 /// DeleteFormula - Remove the given formula from this use's list.
1219 void LSRUse::DeleteFormula(Formula &F) {
1220 if (&F != &Formulae.back())
1221 std::swap(F, Formulae.back());
1222 Formulae.pop_back();
1225 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1226 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1227 // Now that we've filtered out some formulae, recompute the Regs set.
1228 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1230 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1231 E = Formulae.end(); I != E; ++I) {
1232 const Formula &F = *I;
1233 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1234 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1237 // Update the RegTracker.
1238 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1239 E = OldRegs.end(); I != E; ++I)
1240 if (!Regs.count(*I))
1241 RegUses.DropRegister(*I, LUIdx);
1244 void LSRUse::print(raw_ostream &OS) const {
1245 OS << "LSR Use: Kind=";
1247 case Basic: OS << "Basic"; break;
1248 case Special: OS << "Special"; break;
1249 case ICmpZero: OS << "ICmpZero"; break;
1251 OS << "Address of ";
1252 if (AccessTy->isPointerTy())
1253 OS << "pointer"; // the full pointer type could be really verbose
1258 OS << ", Offsets={";
1259 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1260 E = Offsets.end(); I != E; ++I) {
1262 if (llvm::next(I) != E)
1267 if (AllFixupsOutsideLoop)
1268 OS << ", all-fixups-outside-loop";
1270 if (WidestFixupType)
1271 OS << ", widest fixup type: " << *WidestFixupType;
1274 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1275 void LSRUse::dump() const {
1276 print(errs()); errs() << '\n';
1280 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1281 /// be completely folded into the user instruction at isel time. This includes
1282 /// address-mode folding and special icmp tricks.
1283 static bool isLegalUse(const TargetTransformInfo &TTI, LSRUse::KindType Kind,
1284 Type *AccessTy, GlobalValue *BaseGV, int64_t BaseOffset,
1285 bool HasBaseReg, int64_t Scale) {
1287 case LSRUse::Address:
1288 return TTI.isLegalAddressingMode(AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale);
1290 // Otherwise, just guess that reg+reg addressing is legal.
1293 case LSRUse::ICmpZero:
1294 // There's not even a target hook for querying whether it would be legal to
1295 // fold a GV into an ICmp.
1299 // ICmp only has two operands; don't allow more than two non-trivial parts.
1300 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1303 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1304 // putting the scaled register in the other operand of the icmp.
1305 if (Scale != 0 && Scale != -1)
1308 // If we have low-level target information, ask the target if it can fold an
1309 // integer immediate on an icmp.
1310 if (BaseOffset != 0) {
1312 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1313 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1314 // Offs is the ICmp immediate.
1316 // The cast does the right thing with INT64_MIN.
1317 BaseOffset = -(uint64_t)BaseOffset;
1318 return TTI.isLegalICmpImmediate(BaseOffset);
1321 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1325 // Only handle single-register values.
1326 return !BaseGV && Scale == 0 && BaseOffset == 0;
1328 case LSRUse::Special:
1329 // Special case Basic to handle -1 scales.
1330 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1333 llvm_unreachable("Invalid LSRUse Kind!");
1336 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1337 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
1338 GlobalValue *BaseGV, int64_t BaseOffset, bool HasBaseReg,
1340 // Check for overflow.
1341 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1344 MinOffset = (uint64_t)BaseOffset + MinOffset;
1345 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1348 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1350 return isLegalUse(TTI, Kind, AccessTy, BaseGV, MinOffset, HasBaseReg,
1352 isLegalUse(TTI, Kind, AccessTy, BaseGV, MaxOffset, HasBaseReg, Scale);
1355 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1356 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
1358 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1359 F.BaseOffset, F.HasBaseReg, F.Scale);
1362 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1363 LSRUse::KindType Kind, Type *AccessTy,
1364 GlobalValue *BaseGV, int64_t BaseOffset,
1366 // Fast-path: zero is always foldable.
1367 if (BaseOffset == 0 && !BaseGV) return true;
1369 // Conservatively, create an address with an immediate and a
1370 // base and a scale.
1371 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1373 // Canonicalize a scale of 1 to a base register if the formula doesn't
1374 // already have a base register.
1375 if (!HasBaseReg && Scale == 1) {
1380 return isLegalUse(TTI, Kind, AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale);
1383 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1384 ScalarEvolution &SE, int64_t MinOffset,
1385 int64_t MaxOffset, LSRUse::KindType Kind,
1386 Type *AccessTy, const SCEV *S, bool HasBaseReg) {
1387 // Fast-path: zero is always foldable.
1388 if (S->isZero()) return true;
1390 // Conservatively, create an address with an immediate and a
1391 // base and a scale.
1392 int64_t BaseOffset = ExtractImmediate(S, SE);
1393 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1395 // If there's anything else involved, it's not foldable.
1396 if (!S->isZero()) return false;
1398 // Fast-path: zero is always foldable.
1399 if (BaseOffset == 0 && !BaseGV) return true;
1401 // Conservatively, create an address with an immediate and a
1402 // base and a scale.
1403 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1405 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1406 BaseOffset, HasBaseReg, Scale);
1411 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1412 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1413 struct UseMapDenseMapInfo {
1414 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1415 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1418 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1419 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1423 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1424 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1425 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1429 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1430 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1435 /// IVInc - An individual increment in a Chain of IV increments.
1436 /// Relate an IV user to an expression that computes the IV it uses from the IV
1437 /// used by the previous link in the Chain.
1439 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1440 /// original IVOperand. The head of the chain's IVOperand is only valid during
1441 /// chain collection, before LSR replaces IV users. During chain generation,
1442 /// IncExpr can be used to find the new IVOperand that computes the same
1445 Instruction *UserInst;
1447 const SCEV *IncExpr;
1449 IVInc(Instruction *U, Value *O, const SCEV *E):
1450 UserInst(U), IVOperand(O), IncExpr(E) {}
1453 // IVChain - The list of IV increments in program order.
1454 // We typically add the head of a chain without finding subsequent links.
1456 SmallVector<IVInc,1> Incs;
1457 const SCEV *ExprBase;
1459 IVChain() : ExprBase(0) {}
1461 IVChain(const IVInc &Head, const SCEV *Base)
1462 : Incs(1, Head), ExprBase(Base) {}
1464 typedef SmallVectorImpl<IVInc>::const_iterator const_iterator;
1466 // begin - return the first increment in the chain.
1467 const_iterator begin() const {
1468 assert(!Incs.empty());
1469 return llvm::next(Incs.begin());
1471 const_iterator end() const {
1475 // hasIncs - Returns true if this chain contains any increments.
1476 bool hasIncs() const { return Incs.size() >= 2; }
1478 // add - Add an IVInc to the end of this chain.
1479 void add(const IVInc &X) { Incs.push_back(X); }
1481 // tailUserInst - Returns the last UserInst in the chain.
1482 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1484 // isProfitableIncrement - Returns true if IncExpr can be profitably added to
1486 bool isProfitableIncrement(const SCEV *OperExpr,
1487 const SCEV *IncExpr,
1491 /// ChainUsers - Helper for CollectChains to track multiple IV increment uses.
1492 /// Distinguish between FarUsers that definitely cross IV increments and
1493 /// NearUsers that may be used between IV increments.
1495 SmallPtrSet<Instruction*, 4> FarUsers;
1496 SmallPtrSet<Instruction*, 4> NearUsers;
1499 /// LSRInstance - This class holds state for the main loop strength reduction
1503 ScalarEvolution &SE;
1506 const TargetTransformInfo &TTI;
1510 /// IVIncInsertPos - This is the insert position that the current loop's
1511 /// induction variable increment should be placed. In simple loops, this is
1512 /// the latch block's terminator. But in more complicated cases, this is a
1513 /// position which will dominate all the in-loop post-increment users.
1514 Instruction *IVIncInsertPos;
1516 /// Factors - Interesting factors between use strides.
1517 SmallSetVector<int64_t, 8> Factors;
1519 /// Types - Interesting use types, to facilitate truncation reuse.
1520 SmallSetVector<Type *, 4> Types;
1522 /// Fixups - The list of operands which are to be replaced.
1523 SmallVector<LSRFixup, 16> Fixups;
1525 /// Uses - The list of interesting uses.
1526 SmallVector<LSRUse, 16> Uses;
1528 /// RegUses - Track which uses use which register candidates.
1529 RegUseTracker RegUses;
1531 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1532 // have more than a few IV increment chains in a loop. Missing a Chain falls
1533 // back to normal LSR behavior for those uses.
1534 static const unsigned MaxChains = 8;
1536 /// IVChainVec - IV users can form a chain of IV increments.
1537 SmallVector<IVChain, MaxChains> IVChainVec;
1539 /// IVIncSet - IV users that belong to profitable IVChains.
1540 SmallPtrSet<Use*, MaxChains> IVIncSet;
1542 void OptimizeShadowIV();
1543 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1544 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1545 void OptimizeLoopTermCond();
1547 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1548 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1549 void FinalizeChain(IVChain &Chain);
1550 void CollectChains();
1551 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1552 SmallVectorImpl<WeakVH> &DeadInsts);
1554 void CollectInterestingTypesAndFactors();
1555 void CollectFixupsAndInitialFormulae();
1557 LSRFixup &getNewFixup() {
1558 Fixups.push_back(LSRFixup());
1559 return Fixups.back();
1562 // Support for sharing of LSRUses between LSRFixups.
1563 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1565 UseMapDenseMapInfo> UseMapTy;
1568 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1569 LSRUse::KindType Kind, Type *AccessTy);
1571 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1572 LSRUse::KindType Kind,
1575 void DeleteUse(LSRUse &LU, size_t LUIdx);
1577 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1579 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1580 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1581 void CountRegisters(const Formula &F, size_t LUIdx);
1582 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1584 void CollectLoopInvariantFixupsAndFormulae();
1586 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1587 unsigned Depth = 0);
1588 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1589 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1590 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1591 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1592 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1593 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1594 void GenerateCrossUseConstantOffsets();
1595 void GenerateAllReuseFormulae();
1597 void FilterOutUndesirableDedicatedRegisters();
1599 size_t EstimateSearchSpaceComplexity() const;
1600 void NarrowSearchSpaceByDetectingSupersets();
1601 void NarrowSearchSpaceByCollapsingUnrolledCode();
1602 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1603 void NarrowSearchSpaceByPickingWinnerRegs();
1604 void NarrowSearchSpaceUsingHeuristics();
1606 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1608 SmallVectorImpl<const Formula *> &Workspace,
1609 const Cost &CurCost,
1610 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1611 DenseSet<const SCEV *> &VisitedRegs) const;
1612 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1614 BasicBlock::iterator
1615 HoistInsertPosition(BasicBlock::iterator IP,
1616 const SmallVectorImpl<Instruction *> &Inputs) const;
1617 BasicBlock::iterator
1618 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1621 SCEVExpander &Rewriter) const;
1623 Value *Expand(const LSRFixup &LF,
1625 BasicBlock::iterator IP,
1626 SCEVExpander &Rewriter,
1627 SmallVectorImpl<WeakVH> &DeadInsts) const;
1628 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1630 SCEVExpander &Rewriter,
1631 SmallVectorImpl<WeakVH> &DeadInsts,
1633 void Rewrite(const LSRFixup &LF,
1635 SCEVExpander &Rewriter,
1636 SmallVectorImpl<WeakVH> &DeadInsts,
1638 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1642 LSRInstance(Loop *L, Pass *P);
1644 bool getChanged() const { return Changed; }
1646 void print_factors_and_types(raw_ostream &OS) const;
1647 void print_fixups(raw_ostream &OS) const;
1648 void print_uses(raw_ostream &OS) const;
1649 void print(raw_ostream &OS) const;
1655 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1656 /// inside the loop then try to eliminate the cast operation.
1657 void LSRInstance::OptimizeShadowIV() {
1658 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1659 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1662 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1663 UI != E; /* empty */) {
1664 IVUsers::const_iterator CandidateUI = UI;
1666 Instruction *ShadowUse = CandidateUI->getUser();
1667 Type *DestTy = NULL;
1668 bool IsSigned = false;
1670 /* If shadow use is a int->float cast then insert a second IV
1671 to eliminate this cast.
1673 for (unsigned i = 0; i < n; ++i)
1679 for (unsigned i = 0; i < n; ++i, ++d)
1682 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1684 DestTy = UCast->getDestTy();
1686 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1688 DestTy = SCast->getDestTy();
1690 if (!DestTy) continue;
1692 // If target does not support DestTy natively then do not apply
1693 // this transformation.
1694 if (!TTI.isTypeLegal(DestTy)) continue;
1696 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1698 if (PH->getNumIncomingValues() != 2) continue;
1700 Type *SrcTy = PH->getType();
1701 int Mantissa = DestTy->getFPMantissaWidth();
1702 if (Mantissa == -1) continue;
1703 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1706 unsigned Entry, Latch;
1707 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1715 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1716 if (!Init) continue;
1717 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1718 (double)Init->getSExtValue() :
1719 (double)Init->getZExtValue());
1721 BinaryOperator *Incr =
1722 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1723 if (!Incr) continue;
1724 if (Incr->getOpcode() != Instruction::Add
1725 && Incr->getOpcode() != Instruction::Sub)
1728 /* Initialize new IV, double d = 0.0 in above example. */
1729 ConstantInt *C = NULL;
1730 if (Incr->getOperand(0) == PH)
1731 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1732 else if (Incr->getOperand(1) == PH)
1733 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1739 // Ignore negative constants, as the code below doesn't handle them
1740 // correctly. TODO: Remove this restriction.
1741 if (!C->getValue().isStrictlyPositive()) continue;
1743 /* Add new PHINode. */
1744 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1746 /* create new increment. '++d' in above example. */
1747 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1748 BinaryOperator *NewIncr =
1749 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1750 Instruction::FAdd : Instruction::FSub,
1751 NewPH, CFP, "IV.S.next.", Incr);
1753 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1754 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1756 /* Remove cast operation */
1757 ShadowUse->replaceAllUsesWith(NewPH);
1758 ShadowUse->eraseFromParent();
1764 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1765 /// set the IV user and stride information and return true, otherwise return
1767 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1768 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1769 if (UI->getUser() == Cond) {
1770 // NOTE: we could handle setcc instructions with multiple uses here, but
1771 // InstCombine does it as well for simple uses, it's not clear that it
1772 // occurs enough in real life to handle.
1779 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1780 /// a max computation.
1782 /// This is a narrow solution to a specific, but acute, problem. For loops
1788 /// } while (++i < n);
1790 /// the trip count isn't just 'n', because 'n' might not be positive. And
1791 /// unfortunately this can come up even for loops where the user didn't use
1792 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1793 /// will commonly be lowered like this:
1799 /// } while (++i < n);
1802 /// and then it's possible for subsequent optimization to obscure the if
1803 /// test in such a way that indvars can't find it.
1805 /// When indvars can't find the if test in loops like this, it creates a
1806 /// max expression, which allows it to give the loop a canonical
1807 /// induction variable:
1810 /// max = n < 1 ? 1 : n;
1813 /// } while (++i != max);
1815 /// Canonical induction variables are necessary because the loop passes
1816 /// are designed around them. The most obvious example of this is the
1817 /// LoopInfo analysis, which doesn't remember trip count values. It
1818 /// expects to be able to rediscover the trip count each time it is
1819 /// needed, and it does this using a simple analysis that only succeeds if
1820 /// the loop has a canonical induction variable.
1822 /// However, when it comes time to generate code, the maximum operation
1823 /// can be quite costly, especially if it's inside of an outer loop.
1825 /// This function solves this problem by detecting this type of loop and
1826 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1827 /// the instructions for the maximum computation.
1829 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1830 // Check that the loop matches the pattern we're looking for.
1831 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1832 Cond->getPredicate() != CmpInst::ICMP_NE)
1835 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1836 if (!Sel || !Sel->hasOneUse()) return Cond;
1838 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1839 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1841 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1843 // Add one to the backedge-taken count to get the trip count.
1844 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1845 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1847 // Check for a max calculation that matches the pattern. There's no check
1848 // for ICMP_ULE here because the comparison would be with zero, which
1849 // isn't interesting.
1850 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1851 const SCEVNAryExpr *Max = 0;
1852 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1853 Pred = ICmpInst::ICMP_SLE;
1855 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1856 Pred = ICmpInst::ICMP_SLT;
1858 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1859 Pred = ICmpInst::ICMP_ULT;
1866 // To handle a max with more than two operands, this optimization would
1867 // require additional checking and setup.
1868 if (Max->getNumOperands() != 2)
1871 const SCEV *MaxLHS = Max->getOperand(0);
1872 const SCEV *MaxRHS = Max->getOperand(1);
1874 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1875 // for a comparison with 1. For <= and >=, a comparison with zero.
1877 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1880 // Check the relevant induction variable for conformance to
1882 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1883 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1884 if (!AR || !AR->isAffine() ||
1885 AR->getStart() != One ||
1886 AR->getStepRecurrence(SE) != One)
1889 assert(AR->getLoop() == L &&
1890 "Loop condition operand is an addrec in a different loop!");
1892 // Check the right operand of the select, and remember it, as it will
1893 // be used in the new comparison instruction.
1895 if (ICmpInst::isTrueWhenEqual(Pred)) {
1896 // Look for n+1, and grab n.
1897 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1898 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
1899 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1900 NewRHS = BO->getOperand(0);
1901 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1902 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
1903 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1904 NewRHS = BO->getOperand(0);
1907 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1908 NewRHS = Sel->getOperand(1);
1909 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1910 NewRHS = Sel->getOperand(2);
1911 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1912 NewRHS = SU->getValue();
1914 // Max doesn't match expected pattern.
1917 // Determine the new comparison opcode. It may be signed or unsigned,
1918 // and the original comparison may be either equality or inequality.
1919 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1920 Pred = CmpInst::getInversePredicate(Pred);
1922 // Ok, everything looks ok to change the condition into an SLT or SGE and
1923 // delete the max calculation.
1925 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1927 // Delete the max calculation instructions.
1928 Cond->replaceAllUsesWith(NewCond);
1929 CondUse->setUser(NewCond);
1930 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1931 Cond->eraseFromParent();
1932 Sel->eraseFromParent();
1933 if (Cmp->use_empty())
1934 Cmp->eraseFromParent();
1938 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1939 /// postinc iv when possible.
1941 LSRInstance::OptimizeLoopTermCond() {
1942 SmallPtrSet<Instruction *, 4> PostIncs;
1944 BasicBlock *LatchBlock = L->getLoopLatch();
1945 SmallVector<BasicBlock*, 8> ExitingBlocks;
1946 L->getExitingBlocks(ExitingBlocks);
1948 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1949 BasicBlock *ExitingBlock = ExitingBlocks[i];
1951 // Get the terminating condition for the loop if possible. If we
1952 // can, we want to change it to use a post-incremented version of its
1953 // induction variable, to allow coalescing the live ranges for the IV into
1954 // one register value.
1956 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1959 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1960 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1963 // Search IVUsesByStride to find Cond's IVUse if there is one.
1964 IVStrideUse *CondUse = 0;
1965 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1966 if (!FindIVUserForCond(Cond, CondUse))
1969 // If the trip count is computed in terms of a max (due to ScalarEvolution
1970 // being unable to find a sufficient guard, for example), change the loop
1971 // comparison to use SLT or ULT instead of NE.
1972 // One consequence of doing this now is that it disrupts the count-down
1973 // optimization. That's not always a bad thing though, because in such
1974 // cases it may still be worthwhile to avoid a max.
1975 Cond = OptimizeMax(Cond, CondUse);
1977 // If this exiting block dominates the latch block, it may also use
1978 // the post-inc value if it won't be shared with other uses.
1979 // Check for dominance.
1980 if (!DT.dominates(ExitingBlock, LatchBlock))
1983 // Conservatively avoid trying to use the post-inc value in non-latch
1984 // exits if there may be pre-inc users in intervening blocks.
1985 if (LatchBlock != ExitingBlock)
1986 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1987 // Test if the use is reachable from the exiting block. This dominator
1988 // query is a conservative approximation of reachability.
1989 if (&*UI != CondUse &&
1990 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1991 // Conservatively assume there may be reuse if the quotient of their
1992 // strides could be a legal scale.
1993 const SCEV *A = IU.getStride(*CondUse, L);
1994 const SCEV *B = IU.getStride(*UI, L);
1995 if (!A || !B) continue;
1996 if (SE.getTypeSizeInBits(A->getType()) !=
1997 SE.getTypeSizeInBits(B->getType())) {
1998 if (SE.getTypeSizeInBits(A->getType()) >
1999 SE.getTypeSizeInBits(B->getType()))
2000 B = SE.getSignExtendExpr(B, A->getType());
2002 A = SE.getSignExtendExpr(A, B->getType());
2004 if (const SCEVConstant *D =
2005 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2006 const ConstantInt *C = D->getValue();
2007 // Stride of one or negative one can have reuse with non-addresses.
2008 if (C->isOne() || C->isAllOnesValue())
2009 goto decline_post_inc;
2010 // Avoid weird situations.
2011 if (C->getValue().getMinSignedBits() >= 64 ||
2012 C->getValue().isMinSignedValue())
2013 goto decline_post_inc;
2014 // Check for possible scaled-address reuse.
2015 Type *AccessTy = getAccessType(UI->getUser());
2016 int64_t Scale = C->getSExtValue();
2017 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ 0,
2019 /*HasBaseReg=*/ false, Scale))
2020 goto decline_post_inc;
2022 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ 0,
2024 /*HasBaseReg=*/ false, Scale))
2025 goto decline_post_inc;
2029 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2032 // It's possible for the setcc instruction to be anywhere in the loop, and
2033 // possible for it to have multiple users. If it is not immediately before
2034 // the exiting block branch, move it.
2035 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2036 if (Cond->hasOneUse()) {
2037 Cond->moveBefore(TermBr);
2039 // Clone the terminating condition and insert into the loopend.
2040 ICmpInst *OldCond = Cond;
2041 Cond = cast<ICmpInst>(Cond->clone());
2042 Cond->setName(L->getHeader()->getName() + ".termcond");
2043 ExitingBlock->getInstList().insert(TermBr, Cond);
2045 // Clone the IVUse, as the old use still exists!
2046 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2047 TermBr->replaceUsesOfWith(OldCond, Cond);
2051 // If we get to here, we know that we can transform the setcc instruction to
2052 // use the post-incremented version of the IV, allowing us to coalesce the
2053 // live ranges for the IV correctly.
2054 CondUse->transformToPostInc(L);
2057 PostIncs.insert(Cond);
2061 // Determine an insertion point for the loop induction variable increment. It
2062 // must dominate all the post-inc comparisons we just set up, and it must
2063 // dominate the loop latch edge.
2064 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2065 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
2066 E = PostIncs.end(); I != E; ++I) {
2068 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2070 if (BB == (*I)->getParent())
2071 IVIncInsertPos = *I;
2072 else if (BB != IVIncInsertPos->getParent())
2073 IVIncInsertPos = BB->getTerminator();
2077 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
2078 /// at the given offset and other details. If so, update the use and
2081 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
2082 LSRUse::KindType Kind, Type *AccessTy) {
2083 int64_t NewMinOffset = LU.MinOffset;
2084 int64_t NewMaxOffset = LU.MaxOffset;
2085 Type *NewAccessTy = AccessTy;
2087 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2088 // something conservative, however this can pessimize in the case that one of
2089 // the uses will have all its uses outside the loop, for example.
2090 if (LU.Kind != Kind)
2092 // Conservatively assume HasBaseReg is true for now.
2093 if (NewOffset < LU.MinOffset) {
2094 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2095 LU.MaxOffset - NewOffset, HasBaseReg))
2097 NewMinOffset = NewOffset;
2098 } else if (NewOffset > LU.MaxOffset) {
2099 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2100 NewOffset - LU.MinOffset, HasBaseReg))
2102 NewMaxOffset = NewOffset;
2104 // Check for a mismatched access type, and fall back conservatively as needed.
2105 // TODO: Be less conservative when the type is similar and can use the same
2106 // addressing modes.
2107 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
2108 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
2111 LU.MinOffset = NewMinOffset;
2112 LU.MaxOffset = NewMaxOffset;
2113 LU.AccessTy = NewAccessTy;
2114 if (NewOffset != LU.Offsets.back())
2115 LU.Offsets.push_back(NewOffset);
2119 /// getUse - Return an LSRUse index and an offset value for a fixup which
2120 /// needs the given expression, with the given kind and optional access type.
2121 /// Either reuse an existing use or create a new one, as needed.
2122 std::pair<size_t, int64_t>
2123 LSRInstance::getUse(const SCEV *&Expr,
2124 LSRUse::KindType Kind, Type *AccessTy) {
2125 const SCEV *Copy = Expr;
2126 int64_t Offset = ExtractImmediate(Expr, SE);
2128 // Basic uses can't accept any offset, for example.
2129 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2130 Offset, /*HasBaseReg=*/ true)) {
2135 std::pair<UseMapTy::iterator, bool> P =
2136 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
2138 // A use already existed with this base.
2139 size_t LUIdx = P.first->second;
2140 LSRUse &LU = Uses[LUIdx];
2141 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2143 return std::make_pair(LUIdx, Offset);
2146 // Create a new use.
2147 size_t LUIdx = Uses.size();
2148 P.first->second = LUIdx;
2149 Uses.push_back(LSRUse(Kind, AccessTy));
2150 LSRUse &LU = Uses[LUIdx];
2152 // We don't need to track redundant offsets, but we don't need to go out
2153 // of our way here to avoid them.
2154 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
2155 LU.Offsets.push_back(Offset);
2157 LU.MinOffset = Offset;
2158 LU.MaxOffset = Offset;
2159 return std::make_pair(LUIdx, Offset);
2162 /// DeleteUse - Delete the given use from the Uses list.
2163 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2164 if (&LU != &Uses.back())
2165 std::swap(LU, Uses.back());
2169 RegUses.SwapAndDropUse(LUIdx, Uses.size());
2172 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
2173 /// a formula that has the same registers as the given formula.
2175 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2176 const LSRUse &OrigLU) {
2177 // Search all uses for the formula. This could be more clever.
2178 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2179 LSRUse &LU = Uses[LUIdx];
2180 // Check whether this use is close enough to OrigLU, to see whether it's
2181 // worthwhile looking through its formulae.
2182 // Ignore ICmpZero uses because they may contain formulae generated by
2183 // GenerateICmpZeroScales, in which case adding fixup offsets may
2185 if (&LU != &OrigLU &&
2186 LU.Kind != LSRUse::ICmpZero &&
2187 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2188 LU.WidestFixupType == OrigLU.WidestFixupType &&
2189 LU.HasFormulaWithSameRegs(OrigF)) {
2190 // Scan through this use's formulae.
2191 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2192 E = LU.Formulae.end(); I != E; ++I) {
2193 const Formula &F = *I;
2194 // Check to see if this formula has the same registers and symbols
2196 if (F.BaseRegs == OrigF.BaseRegs &&
2197 F.ScaledReg == OrigF.ScaledReg &&
2198 F.BaseGV == OrigF.BaseGV &&
2199 F.Scale == OrigF.Scale &&
2200 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2201 if (F.BaseOffset == 0)
2203 // This is the formula where all the registers and symbols matched;
2204 // there aren't going to be any others. Since we declined it, we
2205 // can skip the rest of the formulae and proceed to the next LSRUse.
2212 // Nothing looked good.
2216 void LSRInstance::CollectInterestingTypesAndFactors() {
2217 SmallSetVector<const SCEV *, 4> Strides;
2219 // Collect interesting types and strides.
2220 SmallVector<const SCEV *, 4> Worklist;
2221 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2222 const SCEV *Expr = IU.getExpr(*UI);
2224 // Collect interesting types.
2225 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2227 // Add strides for mentioned loops.
2228 Worklist.push_back(Expr);
2230 const SCEV *S = Worklist.pop_back_val();
2231 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2232 if (AR->getLoop() == L)
2233 Strides.insert(AR->getStepRecurrence(SE));
2234 Worklist.push_back(AR->getStart());
2235 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2236 Worklist.append(Add->op_begin(), Add->op_end());
2238 } while (!Worklist.empty());
2241 // Compute interesting factors from the set of interesting strides.
2242 for (SmallSetVector<const SCEV *, 4>::const_iterator
2243 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2244 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2245 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
2246 const SCEV *OldStride = *I;
2247 const SCEV *NewStride = *NewStrideIter;
2249 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2250 SE.getTypeSizeInBits(NewStride->getType())) {
2251 if (SE.getTypeSizeInBits(OldStride->getType()) >
2252 SE.getTypeSizeInBits(NewStride->getType()))
2253 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2255 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2257 if (const SCEVConstant *Factor =
2258 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2260 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2261 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2262 } else if (const SCEVConstant *Factor =
2263 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2266 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2267 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2271 // If all uses use the same type, don't bother looking for truncation-based
2273 if (Types.size() == 1)
2276 DEBUG(print_factors_and_types(dbgs()));
2279 /// findIVOperand - Helper for CollectChains that finds an IV operand (computed
2280 /// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped
2281 /// Instructions to IVStrideUses, we could partially skip this.
2282 static User::op_iterator
2283 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2284 Loop *L, ScalarEvolution &SE) {
2285 for(; OI != OE; ++OI) {
2286 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2287 if (!SE.isSCEVable(Oper->getType()))
2290 if (const SCEVAddRecExpr *AR =
2291 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2292 if (AR->getLoop() == L)
2300 /// getWideOperand - IVChain logic must consistenctly peek base TruncInst
2301 /// operands, so wrap it in a convenient helper.
2302 static Value *getWideOperand(Value *Oper) {
2303 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2304 return Trunc->getOperand(0);
2308 /// isCompatibleIVType - Return true if we allow an IV chain to include both
2310 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2311 Type *LType = LVal->getType();
2312 Type *RType = RVal->getType();
2313 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy());
2316 /// getExprBase - Return an approximation of this SCEV expression's "base", or
2317 /// NULL for any constant. Returning the expression itself is
2318 /// conservative. Returning a deeper subexpression is more precise and valid as
2319 /// long as it isn't less complex than another subexpression. For expressions
2320 /// involving multiple unscaled values, we need to return the pointer-type
2321 /// SCEVUnknown. This avoids forming chains across objects, such as:
2322 /// PrevOper==a[i], IVOper==b[i], IVInc==b-a.
2324 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2325 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2326 static const SCEV *getExprBase(const SCEV *S) {
2327 switch (S->getSCEVType()) {
2328 default: // uncluding scUnknown.
2333 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2335 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2337 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2339 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2340 // there's nothing more complex.
2341 // FIXME: not sure if we want to recognize negation.
2342 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2343 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2344 E(Add->op_begin()); I != E; ++I) {
2345 const SCEV *SubExpr = *I;
2346 if (SubExpr->getSCEVType() == scAddExpr)
2347 return getExprBase(SubExpr);
2349 if (SubExpr->getSCEVType() != scMulExpr)
2352 return S; // all operands are scaled, be conservative.
2355 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2359 /// Return true if the chain increment is profitable to expand into a loop
2360 /// invariant value, which may require its own register. A profitable chain
2361 /// increment will be an offset relative to the same base. We allow such offsets
2362 /// to potentially be used as chain increment as long as it's not obviously
2363 /// expensive to expand using real instructions.
2364 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2365 const SCEV *IncExpr,
2366 ScalarEvolution &SE) {
2367 // Aggressively form chains when -stress-ivchain.
2371 // Do not replace a constant offset from IV head with a nonconstant IV
2373 if (!isa<SCEVConstant>(IncExpr)) {
2374 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2375 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2379 SmallPtrSet<const SCEV*, 8> Processed;
2380 return !isHighCostExpansion(IncExpr, Processed, SE);
2383 /// Return true if the number of registers needed for the chain is estimated to
2384 /// be less than the number required for the individual IV users. First prohibit
2385 /// any IV users that keep the IV live across increments (the Users set should
2386 /// be empty). Next count the number and type of increments in the chain.
2388 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2389 /// effectively use postinc addressing modes. Only consider it profitable it the
2390 /// increments can be computed in fewer registers when chained.
2392 /// TODO: Consider IVInc free if it's already used in another chains.
2394 isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users,
2395 ScalarEvolution &SE, const TargetTransformInfo &TTI) {
2399 if (!Chain.hasIncs())
2402 if (!Users.empty()) {
2403 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2404 for (SmallPtrSet<Instruction*, 4>::const_iterator I = Users.begin(),
2405 E = Users.end(); I != E; ++I) {
2406 dbgs() << " " << **I << "\n";
2410 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2412 // The chain itself may require a register, so intialize cost to 1.
2415 // A complete chain likely eliminates the need for keeping the original IV in
2416 // a register. LSR does not currently know how to form a complete chain unless
2417 // the header phi already exists.
2418 if (isa<PHINode>(Chain.tailUserInst())
2419 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2422 const SCEV *LastIncExpr = 0;
2423 unsigned NumConstIncrements = 0;
2424 unsigned NumVarIncrements = 0;
2425 unsigned NumReusedIncrements = 0;
2426 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2429 if (I->IncExpr->isZero())
2432 // Incrementing by zero or some constant is neutral. We assume constants can
2433 // be folded into an addressing mode or an add's immediate operand.
2434 if (isa<SCEVConstant>(I->IncExpr)) {
2435 ++NumConstIncrements;
2439 if (I->IncExpr == LastIncExpr)
2440 ++NumReusedIncrements;
2444 LastIncExpr = I->IncExpr;
2446 // An IV chain with a single increment is handled by LSR's postinc
2447 // uses. However, a chain with multiple increments requires keeping the IV's
2448 // value live longer than it needs to be if chained.
2449 if (NumConstIncrements > 1)
2452 // Materializing increment expressions in the preheader that didn't exist in
2453 // the original code may cost a register. For example, sign-extended array
2454 // indices can produce ridiculous increments like this:
2455 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2456 cost += NumVarIncrements;
2458 // Reusing variable increments likely saves a register to hold the multiple of
2460 cost -= NumReusedIncrements;
2462 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2468 /// ChainInstruction - Add this IV user to an existing chain or make it the head
2470 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2471 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2472 // When IVs are used as types of varying widths, they are generally converted
2473 // to a wider type with some uses remaining narrow under a (free) trunc.
2474 Value *const NextIV = getWideOperand(IVOper);
2475 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2476 const SCEV *const OperExprBase = getExprBase(OperExpr);
2478 // Visit all existing chains. Check if its IVOper can be computed as a
2479 // profitable loop invariant increment from the last link in the Chain.
2480 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2481 const SCEV *LastIncExpr = 0;
2482 for (; ChainIdx < NChains; ++ChainIdx) {
2483 IVChain &Chain = IVChainVec[ChainIdx];
2485 // Prune the solution space aggressively by checking that both IV operands
2486 // are expressions that operate on the same unscaled SCEVUnknown. This
2487 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2488 // first avoids creating extra SCEV expressions.
2489 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2492 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2493 if (!isCompatibleIVType(PrevIV, NextIV))
2496 // A phi node terminates a chain.
2497 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2500 // The increment must be loop-invariant so it can be kept in a register.
2501 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2502 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2503 if (!SE.isLoopInvariant(IncExpr, L))
2506 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2507 LastIncExpr = IncExpr;
2511 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2512 // bother for phi nodes, because they must be last in the chain.
2513 if (ChainIdx == NChains) {
2514 if (isa<PHINode>(UserInst))
2516 if (NChains >= MaxChains && !StressIVChain) {
2517 DEBUG(dbgs() << "IV Chain Limit\n");
2520 LastIncExpr = OperExpr;
2521 // IVUsers may have skipped over sign/zero extensions. We don't currently
2522 // attempt to form chains involving extensions unless they can be hoisted
2523 // into this loop's AddRec.
2524 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2527 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2529 ChainUsersVec.resize(NChains);
2530 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2531 << ") IV=" << *LastIncExpr << "\n");
2533 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2534 << ") IV+" << *LastIncExpr << "\n");
2535 // Add this IV user to the end of the chain.
2536 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2538 IVChain &Chain = IVChainVec[ChainIdx];
2540 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2541 // This chain's NearUsers become FarUsers.
2542 if (!LastIncExpr->isZero()) {
2543 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2548 // All other uses of IVOperand become near uses of the chain.
2549 // We currently ignore intermediate values within SCEV expressions, assuming
2550 // they will eventually be used be the current chain, or can be computed
2551 // from one of the chain increments. To be more precise we could
2552 // transitively follow its user and only add leaf IV users to the set.
2553 for (Value::use_iterator UseIter = IVOper->use_begin(),
2554 UseEnd = IVOper->use_end(); UseIter != UseEnd; ++UseIter) {
2555 Instruction *OtherUse = dyn_cast<Instruction>(*UseIter);
2558 // Uses in the chain will no longer be uses if the chain is formed.
2559 // Include the head of the chain in this iteration (not Chain.begin()).
2560 IVChain::const_iterator IncIter = Chain.Incs.begin();
2561 IVChain::const_iterator IncEnd = Chain.Incs.end();
2562 for( ; IncIter != IncEnd; ++IncIter) {
2563 if (IncIter->UserInst == OtherUse)
2566 if (IncIter != IncEnd)
2569 if (SE.isSCEVable(OtherUse->getType())
2570 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2571 && IU.isIVUserOrOperand(OtherUse)) {
2574 NearUsers.insert(OtherUse);
2577 // Since this user is part of the chain, it's no longer considered a use
2579 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2582 /// CollectChains - Populate the vector of Chains.
2584 /// This decreases ILP at the architecture level. Targets with ample registers,
2585 /// multiple memory ports, and no register renaming probably don't want
2586 /// this. However, such targets should probably disable LSR altogether.
2588 /// The job of LSR is to make a reasonable choice of induction variables across
2589 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2590 /// ILP *within the loop* if the target wants it.
2592 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2593 /// will not reorder memory operations, it will recognize this as a chain, but
2594 /// will generate redundant IV increments. Ideally this would be corrected later
2595 /// by a smart scheduler:
2601 /// TODO: Walk the entire domtree within this loop, not just the path to the
2602 /// loop latch. This will discover chains on side paths, but requires
2603 /// maintaining multiple copies of the Chains state.
2604 void LSRInstance::CollectChains() {
2605 DEBUG(dbgs() << "Collecting IV Chains.\n");
2606 SmallVector<ChainUsers, 8> ChainUsersVec;
2608 SmallVector<BasicBlock *,8> LatchPath;
2609 BasicBlock *LoopHeader = L->getHeader();
2610 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2611 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2612 LatchPath.push_back(Rung->getBlock());
2614 LatchPath.push_back(LoopHeader);
2616 // Walk the instruction stream from the loop header to the loop latch.
2617 for (SmallVectorImpl<BasicBlock *>::reverse_iterator
2618 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend();
2619 BBIter != BBEnd; ++BBIter) {
2620 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end();
2622 // Skip instructions that weren't seen by IVUsers analysis.
2623 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I))
2626 // Ignore users that are part of a SCEV expression. This way we only
2627 // consider leaf IV Users. This effectively rediscovers a portion of
2628 // IVUsers analysis but in program order this time.
2629 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I)))
2632 // Remove this instruction from any NearUsers set it may be in.
2633 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2634 ChainIdx < NChains; ++ChainIdx) {
2635 ChainUsersVec[ChainIdx].NearUsers.erase(I);
2637 // Search for operands that can be chained.
2638 SmallPtrSet<Instruction*, 4> UniqueOperands;
2639 User::op_iterator IVOpEnd = I->op_end();
2640 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE);
2641 while (IVOpIter != IVOpEnd) {
2642 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2643 if (UniqueOperands.insert(IVOpInst))
2644 ChainInstruction(I, IVOpInst, ChainUsersVec);
2645 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE);
2647 } // Continue walking down the instructions.
2648 } // Continue walking down the domtree.
2649 // Visit phi backedges to determine if the chain can generate the IV postinc.
2650 for (BasicBlock::iterator I = L->getHeader()->begin();
2651 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2652 if (!SE.isSCEVable(PN->getType()))
2656 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2658 ChainInstruction(PN, IncV, ChainUsersVec);
2660 // Remove any unprofitable chains.
2661 unsigned ChainIdx = 0;
2662 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2663 UsersIdx < NChains; ++UsersIdx) {
2664 if (!isProfitableChain(IVChainVec[UsersIdx],
2665 ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
2667 // Preserve the chain at UsesIdx.
2668 if (ChainIdx != UsersIdx)
2669 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2670 FinalizeChain(IVChainVec[ChainIdx]);
2673 IVChainVec.resize(ChainIdx);
2676 void LSRInstance::FinalizeChain(IVChain &Chain) {
2677 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2678 DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
2680 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2682 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n");
2683 User::op_iterator UseI =
2684 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand);
2685 assert(UseI != I->UserInst->op_end() && "cannot find IV operand");
2686 IVIncSet.insert(UseI);
2690 /// Return true if the IVInc can be folded into an addressing mode.
2691 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
2692 Value *Operand, const TargetTransformInfo &TTI) {
2693 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
2694 if (!IncConst || !isAddressUse(UserInst, Operand))
2697 if (IncConst->getValue()->getValue().getMinSignedBits() > 64)
2700 int64_t IncOffset = IncConst->getValue()->getSExtValue();
2701 if (!isAlwaysFoldable(TTI, LSRUse::Address,
2702 getAccessType(UserInst), /*BaseGV=*/ 0,
2703 IncOffset, /*HaseBaseReg=*/ false))
2709 /// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to
2710 /// materialize the IV user's operand from the previous IV user's operand.
2711 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
2712 SmallVectorImpl<WeakVH> &DeadInsts) {
2713 // Find the new IVOperand for the head of the chain. It may have been replaced
2715 const IVInc &Head = Chain.Incs[0];
2716 User::op_iterator IVOpEnd = Head.UserInst->op_end();
2717 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
2718 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
2721 while (IVOpIter != IVOpEnd) {
2722 IVSrc = getWideOperand(*IVOpIter);
2724 // If this operand computes the expression that the chain needs, we may use
2725 // it. (Check this after setting IVSrc which is used below.)
2727 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
2728 // narrow for the chain, so we can no longer use it. We do allow using a
2729 // wider phi, assuming the LSR checked for free truncation. In that case we
2730 // should already have a truncate on this operand such that
2731 // getSCEV(IVSrc) == IncExpr.
2732 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
2733 || SE.getSCEV(IVSrc) == Head.IncExpr) {
2736 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE);
2738 if (IVOpIter == IVOpEnd) {
2739 // Gracefully give up on this chain.
2740 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
2744 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
2745 Type *IVTy = IVSrc->getType();
2746 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
2747 const SCEV *LeftOverExpr = 0;
2748 for (IVChain::const_iterator IncI = Chain.begin(),
2749 IncE = Chain.end(); IncI != IncE; ++IncI) {
2751 Instruction *InsertPt = IncI->UserInst;
2752 if (isa<PHINode>(InsertPt))
2753 InsertPt = L->getLoopLatch()->getTerminator();
2755 // IVOper will replace the current IV User's operand. IVSrc is the IV
2756 // value currently held in a register.
2757 Value *IVOper = IVSrc;
2758 if (!IncI->IncExpr->isZero()) {
2759 // IncExpr was the result of subtraction of two narrow values, so must
2761 const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy);
2762 LeftOverExpr = LeftOverExpr ?
2763 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
2765 if (LeftOverExpr && !LeftOverExpr->isZero()) {
2766 // Expand the IV increment.
2767 Rewriter.clearPostInc();
2768 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
2769 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
2770 SE.getUnknown(IncV));
2771 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
2773 // If an IV increment can't be folded, use it as the next IV value.
2774 if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand,
2776 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
2781 Type *OperTy = IncI->IVOperand->getType();
2782 if (IVTy != OperTy) {
2783 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
2784 "cannot extend a chained IV");
2785 IRBuilder<> Builder(InsertPt);
2786 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
2788 IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper);
2789 DeadInsts.push_back(IncI->IVOperand);
2791 // If LSR created a new, wider phi, we may also replace its postinc. We only
2792 // do this if we also found a wide value for the head of the chain.
2793 if (isa<PHINode>(Chain.tailUserInst())) {
2794 for (BasicBlock::iterator I = L->getHeader()->begin();
2795 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
2796 if (!isCompatibleIVType(Phi, IVSrc))
2798 Instruction *PostIncV = dyn_cast<Instruction>(
2799 Phi->getIncomingValueForBlock(L->getLoopLatch()));
2800 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
2802 Value *IVOper = IVSrc;
2803 Type *PostIncTy = PostIncV->getType();
2804 if (IVTy != PostIncTy) {
2805 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
2806 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
2807 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
2808 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
2810 Phi->replaceUsesOfWith(PostIncV, IVOper);
2811 DeadInsts.push_back(PostIncV);
2816 void LSRInstance::CollectFixupsAndInitialFormulae() {
2817 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2818 Instruction *UserInst = UI->getUser();
2819 // Skip IV users that are part of profitable IV Chains.
2820 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(),
2821 UI->getOperandValToReplace());
2822 assert(UseI != UserInst->op_end() && "cannot find IV operand");
2823 if (IVIncSet.count(UseI))
2827 LSRFixup &LF = getNewFixup();
2828 LF.UserInst = UserInst;
2829 LF.OperandValToReplace = UI->getOperandValToReplace();
2830 LF.PostIncLoops = UI->getPostIncLoops();
2832 LSRUse::KindType Kind = LSRUse::Basic;
2834 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2835 Kind = LSRUse::Address;
2836 AccessTy = getAccessType(LF.UserInst);
2839 const SCEV *S = IU.getExpr(*UI);
2841 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2842 // (N - i == 0), and this allows (N - i) to be the expression that we work
2843 // with rather than just N or i, so we can consider the register
2844 // requirements for both N and i at the same time. Limiting this code to
2845 // equality icmps is not a problem because all interesting loops use
2846 // equality icmps, thanks to IndVarSimplify.
2847 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2848 if (CI->isEquality()) {
2849 // Swap the operands if needed to put the OperandValToReplace on the
2850 // left, for consistency.
2851 Value *NV = CI->getOperand(1);
2852 if (NV == LF.OperandValToReplace) {
2853 CI->setOperand(1, CI->getOperand(0));
2854 CI->setOperand(0, NV);
2855 NV = CI->getOperand(1);
2859 // x == y --> x - y == 0
2860 const SCEV *N = SE.getSCEV(NV);
2861 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N)) {
2862 // S is normalized, so normalize N before folding it into S
2863 // to keep the result normalized.
2864 N = TransformForPostIncUse(Normalize, N, CI, 0,
2865 LF.PostIncLoops, SE, DT);
2866 Kind = LSRUse::ICmpZero;
2867 S = SE.getMinusSCEV(N, S);
2870 // -1 and the negations of all interesting strides (except the negation
2871 // of -1) are now also interesting.
2872 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2873 if (Factors[i] != -1)
2874 Factors.insert(-(uint64_t)Factors[i]);
2878 // Set up the initial formula for this use.
2879 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2881 LF.Offset = P.second;
2882 LSRUse &LU = Uses[LF.LUIdx];
2883 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2884 if (!LU.WidestFixupType ||
2885 SE.getTypeSizeInBits(LU.WidestFixupType) <
2886 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2887 LU.WidestFixupType = LF.OperandValToReplace->getType();
2889 // If this is the first use of this LSRUse, give it a formula.
2890 if (LU.Formulae.empty()) {
2891 InsertInitialFormula(S, LU, LF.LUIdx);
2892 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2896 DEBUG(print_fixups(dbgs()));
2899 /// InsertInitialFormula - Insert a formula for the given expression into
2900 /// the given use, separating out loop-variant portions from loop-invariant
2901 /// and loop-computable portions.
2903 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2905 F.InitialMatch(S, L, SE);
2906 bool Inserted = InsertFormula(LU, LUIdx, F);
2907 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2910 /// InsertSupplementalFormula - Insert a simple single-register formula for
2911 /// the given expression into the given use.
2913 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2914 LSRUse &LU, size_t LUIdx) {
2916 F.BaseRegs.push_back(S);
2917 F.HasBaseReg = true;
2918 bool Inserted = InsertFormula(LU, LUIdx, F);
2919 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2922 /// CountRegisters - Note which registers are used by the given formula,
2923 /// updating RegUses.
2924 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2926 RegUses.CountRegister(F.ScaledReg, LUIdx);
2927 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2928 E = F.BaseRegs.end(); I != E; ++I)
2929 RegUses.CountRegister(*I, LUIdx);
2932 /// InsertFormula - If the given formula has not yet been inserted, add it to
2933 /// the list, and return true. Return false otherwise.
2934 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2935 if (!LU.InsertFormula(F))
2938 CountRegisters(F, LUIdx);
2942 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2943 /// loop-invariant values which we're tracking. These other uses will pin these
2944 /// values in registers, making them less profitable for elimination.
2945 /// TODO: This currently misses non-constant addrec step registers.
2946 /// TODO: Should this give more weight to users inside the loop?
2948 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2949 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2950 SmallPtrSet<const SCEV *, 8> Inserted;
2952 while (!Worklist.empty()) {
2953 const SCEV *S = Worklist.pop_back_val();
2955 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2956 Worklist.append(N->op_begin(), N->op_end());
2957 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2958 Worklist.push_back(C->getOperand());
2959 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2960 Worklist.push_back(D->getLHS());
2961 Worklist.push_back(D->getRHS());
2962 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2963 if (!Inserted.insert(U)) continue;
2964 const Value *V = U->getValue();
2965 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2966 // Look for instructions defined outside the loop.
2967 if (L->contains(Inst)) continue;
2968 } else if (isa<UndefValue>(V))
2969 // Undef doesn't have a live range, so it doesn't matter.
2971 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2973 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2974 // Ignore non-instructions.
2977 // Ignore instructions in other functions (as can happen with
2979 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2981 // Ignore instructions not dominated by the loop.
2982 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2983 UserInst->getParent() :
2984 cast<PHINode>(UserInst)->getIncomingBlock(
2985 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2986 if (!DT.dominates(L->getHeader(), UseBB))
2988 // Ignore uses which are part of other SCEV expressions, to avoid
2989 // analyzing them multiple times.
2990 if (SE.isSCEVable(UserInst->getType())) {
2991 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2992 // If the user is a no-op, look through to its uses.
2993 if (!isa<SCEVUnknown>(UserS))
2997 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3001 // Ignore icmp instructions which are already being analyzed.
3002 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3003 unsigned OtherIdx = !UI.getOperandNo();
3004 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3005 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3009 LSRFixup &LF = getNewFixup();
3010 LF.UserInst = const_cast<Instruction *>(UserInst);
3011 LF.OperandValToReplace = UI.getUse();
3012 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
3014 LF.Offset = P.second;
3015 LSRUse &LU = Uses[LF.LUIdx];
3016 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3017 if (!LU.WidestFixupType ||
3018 SE.getTypeSizeInBits(LU.WidestFixupType) <
3019 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3020 LU.WidestFixupType = LF.OperandValToReplace->getType();
3021 InsertSupplementalFormula(U, LU, LF.LUIdx);
3022 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3029 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
3030 /// separate registers. If C is non-null, multiply each subexpression by C.
3032 /// Return remainder expression after factoring the subexpressions captured by
3033 /// Ops. If Ops is complete, return NULL.
3034 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3035 SmallVectorImpl<const SCEV *> &Ops,
3037 ScalarEvolution &SE,
3038 unsigned Depth = 0) {
3039 // Arbitrarily cap recursion to protect compile time.
3043 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3044 // Break out add operands.
3045 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
3047 const SCEV *Remainder = CollectSubexprs(*I, C, Ops, L, SE, Depth+1);
3049 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3052 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3053 // Split a non-zero base out of an addrec.
3054 if (AR->getStart()->isZero())
3057 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3058 C, Ops, L, SE, Depth+1);
3059 // Split the non-zero AddRec unless it is part of a nested recurrence that
3060 // does not pertain to this loop.
3061 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3062 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3065 if (Remainder != AR->getStart()) {
3067 Remainder = SE.getConstant(AR->getType(), 0);
3068 return SE.getAddRecExpr(Remainder,
3069 AR->getStepRecurrence(SE),
3071 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3074 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3075 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3076 if (Mul->getNumOperands() != 2)
3078 if (const SCEVConstant *Op0 =
3079 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3080 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3081 const SCEV *Remainder =
3082 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3084 Ops.push_back(SE.getMulExpr(C, Remainder));
3091 /// GenerateReassociations - Split out subexpressions from adds and the bases of
3093 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3096 // Arbitrarily cap recursion to protect compile time.
3097 if (Depth >= 3) return;
3099 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3100 const SCEV *BaseReg = Base.BaseRegs[i];
3102 SmallVector<const SCEV *, 8> AddOps;
3103 const SCEV *Remainder = CollectSubexprs(BaseReg, 0, AddOps, L, SE);
3105 AddOps.push_back(Remainder);
3107 if (AddOps.size() == 1) continue;
3109 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3110 JE = AddOps.end(); J != JE; ++J) {
3112 // Loop-variant "unknown" values are uninteresting; we won't be able to
3113 // do anything meaningful with them.
3114 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3117 // Don't pull a constant into a register if the constant could be folded
3118 // into an immediate field.
3119 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3120 LU.AccessTy, *J, Base.getNumRegs() > 1))
3123 // Collect all operands except *J.
3124 SmallVector<const SCEV *, 8> InnerAddOps
3125 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3127 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3129 // Don't leave just a constant behind in a register if the constant could
3130 // be folded into an immediate field.
3131 if (InnerAddOps.size() == 1 &&
3132 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3133 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3136 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3137 if (InnerSum->isZero())
3141 // Add the remaining pieces of the add back into the new formula.
3142 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3144 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3145 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3146 InnerSumSC->getValue()->getZExtValue())) {
3147 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3148 InnerSumSC->getValue()->getZExtValue();
3149 F.BaseRegs.erase(F.BaseRegs.begin() + i);
3151 F.BaseRegs[i] = InnerSum;
3153 // Add J as its own register, or an unfolded immediate.
3154 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3155 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3156 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3157 SC->getValue()->getZExtValue()))
3158 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3159 SC->getValue()->getZExtValue();
3161 F.BaseRegs.push_back(*J);
3163 if (InsertFormula(LU, LUIdx, F))
3164 // If that formula hadn't been seen before, recurse to find more like
3166 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
3171 /// GenerateCombinations - Generate a formula consisting of all of the
3172 /// loop-dominating registers added into a single register.
3173 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3175 // This method is only interesting on a plurality of registers.
3176 if (Base.BaseRegs.size() <= 1) return;
3180 SmallVector<const SCEV *, 4> Ops;
3181 for (SmallVectorImpl<const SCEV *>::const_iterator
3182 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
3183 const SCEV *BaseReg = *I;
3184 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3185 !SE.hasComputableLoopEvolution(BaseReg, L))
3186 Ops.push_back(BaseReg);
3188 F.BaseRegs.push_back(BaseReg);
3190 if (Ops.size() > 1) {
3191 const SCEV *Sum = SE.getAddExpr(Ops);
3192 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3193 // opportunity to fold something. For now, just ignore such cases
3194 // rather than proceed with zero in a register.
3195 if (!Sum->isZero()) {
3196 F.BaseRegs.push_back(Sum);
3197 (void)InsertFormula(LU, LUIdx, F);
3202 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
3203 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3205 // We can't add a symbolic offset if the address already contains one.
3206 if (Base.BaseGV) return;
3208 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3209 const SCEV *G = Base.BaseRegs[i];
3210 GlobalValue *GV = ExtractSymbol(G, SE);
3211 if (G->isZero() || !GV)
3215 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3218 (void)InsertFormula(LU, LUIdx, F);
3222 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3223 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3225 // TODO: For now, just add the min and max offset, because it usually isn't
3226 // worthwhile looking at everything inbetween.
3227 SmallVector<int64_t, 2> Worklist;
3228 Worklist.push_back(LU.MinOffset);
3229 if (LU.MaxOffset != LU.MinOffset)
3230 Worklist.push_back(LU.MaxOffset);
3232 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3233 const SCEV *G = Base.BaseRegs[i];
3235 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
3236 E = Worklist.end(); I != E; ++I) {
3238 F.BaseOffset = (uint64_t)Base.BaseOffset - *I;
3239 if (isLegalUse(TTI, LU.MinOffset - *I, LU.MaxOffset - *I, LU.Kind,
3241 // Add the offset to the base register.
3242 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
3243 // If it cancelled out, drop the base register, otherwise update it.
3244 if (NewG->isZero()) {
3245 std::swap(F.BaseRegs[i], F.BaseRegs.back());
3246 F.BaseRegs.pop_back();
3248 F.BaseRegs[i] = NewG;
3250 (void)InsertFormula(LU, LUIdx, F);
3254 int64_t Imm = ExtractImmediate(G, SE);
3255 if (G->isZero() || Imm == 0)
3258 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3259 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3262 (void)InsertFormula(LU, LUIdx, F);
3266 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
3267 /// the comparison. For example, x == y -> x*c == y*c.
3268 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3270 if (LU.Kind != LSRUse::ICmpZero) return;
3272 // Determine the integer type for the base formula.
3273 Type *IntTy = Base.getType();
3275 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3277 // Don't do this if there is more than one offset.
3278 if (LU.MinOffset != LU.MaxOffset) return;
3280 assert(!Base.BaseGV && "ICmpZero use is not legal!");
3282 // Check each interesting stride.
3283 for (SmallSetVector<int64_t, 8>::const_iterator
3284 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3285 int64_t Factor = *I;
3287 // Check that the multiplication doesn't overflow.
3288 if (Base.BaseOffset == INT64_MIN && Factor == -1)
3290 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3291 if (NewBaseOffset / Factor != Base.BaseOffset)
3294 // Check that multiplying with the use offset doesn't overflow.
3295 int64_t Offset = LU.MinOffset;
3296 if (Offset == INT64_MIN && Factor == -1)
3298 Offset = (uint64_t)Offset * Factor;
3299 if (Offset / Factor != LU.MinOffset)
3303 F.BaseOffset = NewBaseOffset;
3305 // Check that this scale is legal.
3306 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3309 // Compensate for the use having MinOffset built into it.
3310 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3312 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3314 // Check that multiplying with each base register doesn't overflow.
3315 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3316 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3317 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3321 // Check that multiplying with the scaled register doesn't overflow.
3323 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3324 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3328 // Check that multiplying with the unfolded offset doesn't overflow.
3329 if (F.UnfoldedOffset != 0) {
3330 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3332 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3333 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3337 // If we make it here and it's legal, add it.
3338 (void)InsertFormula(LU, LUIdx, F);
3343 /// GenerateScales - Generate stride factor reuse formulae by making use of
3344 /// scaled-offset address modes, for example.
3345 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3346 // Determine the integer type for the base formula.
3347 Type *IntTy = Base.getType();
3350 // If this Formula already has a scaled register, we can't add another one.
3351 if (Base.Scale != 0) return;
3353 // Check each interesting stride.
3354 for (SmallSetVector<int64_t, 8>::const_iterator
3355 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3356 int64_t Factor = *I;
3358 Base.Scale = Factor;
3359 Base.HasBaseReg = Base.BaseRegs.size() > 1;
3360 // Check whether this scale is going to be legal.
3361 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3363 // As a special-case, handle special out-of-loop Basic users specially.
3364 // TODO: Reconsider this special case.
3365 if (LU.Kind == LSRUse::Basic &&
3366 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3367 LU.AccessTy, Base) &&
3368 LU.AllFixupsOutsideLoop)
3369 LU.Kind = LSRUse::Special;
3373 // For an ICmpZero, negating a solitary base register won't lead to
3375 if (LU.Kind == LSRUse::ICmpZero &&
3376 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3378 // For each addrec base reg, apply the scale, if possible.
3379 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3380 if (const SCEVAddRecExpr *AR =
3381 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
3382 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3383 if (FactorS->isZero())
3385 // Divide out the factor, ignoring high bits, since we'll be
3386 // scaling the value back up in the end.
3387 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3388 // TODO: This could be optimized to avoid all the copying.
3390 F.ScaledReg = Quotient;
3391 F.DeleteBaseReg(F.BaseRegs[i]);
3392 (void)InsertFormula(LU, LUIdx, F);
3398 /// GenerateTruncates - Generate reuse formulae from different IV types.
3399 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3400 // Don't bother truncating symbolic values.
3401 if (Base.BaseGV) return;
3403 // Determine the integer type for the base formula.
3404 Type *DstTy = Base.getType();
3406 DstTy = SE.getEffectiveSCEVType(DstTy);
3408 for (SmallSetVector<Type *, 4>::const_iterator
3409 I = Types.begin(), E = Types.end(); I != E; ++I) {
3411 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
3414 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
3415 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
3416 JE = F.BaseRegs.end(); J != JE; ++J)
3417 *J = SE.getAnyExtendExpr(*J, SrcTy);
3419 // TODO: This assumes we've done basic processing on all uses and
3420 // have an idea what the register usage is.
3421 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3424 (void)InsertFormula(LU, LUIdx, F);
3431 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
3432 /// defer modifications so that the search phase doesn't have to worry about
3433 /// the data structures moving underneath it.
3437 const SCEV *OrigReg;
3439 WorkItem(size_t LI, int64_t I, const SCEV *R)
3440 : LUIdx(LI), Imm(I), OrigReg(R) {}
3442 void print(raw_ostream &OS) const;
3448 void WorkItem::print(raw_ostream &OS) const {
3449 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3450 << " , add offset " << Imm;
3453 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3454 void WorkItem::dump() const {
3455 print(errs()); errs() << '\n';
3459 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
3460 /// distance apart and try to form reuse opportunities between them.
3461 void LSRInstance::GenerateCrossUseConstantOffsets() {
3462 // Group the registers by their value without any added constant offset.
3463 typedef std::map<int64_t, const SCEV *> ImmMapTy;
3464 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
3466 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3467 SmallVector<const SCEV *, 8> Sequence;
3468 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3470 const SCEV *Reg = *I;
3471 int64_t Imm = ExtractImmediate(Reg, SE);
3472 std::pair<RegMapTy::iterator, bool> Pair =
3473 Map.insert(std::make_pair(Reg, ImmMapTy()));
3475 Sequence.push_back(Reg);
3476 Pair.first->second.insert(std::make_pair(Imm, *I));
3477 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
3480 // Now examine each set of registers with the same base value. Build up
3481 // a list of work to do and do the work in a separate step so that we're
3482 // not adding formulae and register counts while we're searching.
3483 SmallVector<WorkItem, 32> WorkItems;
3484 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3485 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
3486 E = Sequence.end(); I != E; ++I) {
3487 const SCEV *Reg = *I;
3488 const ImmMapTy &Imms = Map.find(Reg)->second;
3490 // It's not worthwhile looking for reuse if there's only one offset.
3491 if (Imms.size() == 1)
3494 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3495 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3497 dbgs() << ' ' << J->first;
3500 // Examine each offset.
3501 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3503 const SCEV *OrigReg = J->second;
3505 int64_t JImm = J->first;
3506 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3508 if (!isa<SCEVConstant>(OrigReg) &&
3509 UsedByIndicesMap[Reg].count() == 1) {
3510 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3514 // Conservatively examine offsets between this orig reg a few selected
3516 ImmMapTy::const_iterator OtherImms[] = {
3517 Imms.begin(), prior(Imms.end()),
3518 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
3520 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3521 ImmMapTy::const_iterator M = OtherImms[i];
3522 if (M == J || M == JE) continue;
3524 // Compute the difference between the two.
3525 int64_t Imm = (uint64_t)JImm - M->first;
3526 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
3527 LUIdx = UsedByIndices.find_next(LUIdx))
3528 // Make a memo of this use, offset, and register tuple.
3529 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
3530 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3537 UsedByIndicesMap.clear();
3538 UniqueItems.clear();
3540 // Now iterate through the worklist and add new formulae.
3541 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
3542 E = WorkItems.end(); I != E; ++I) {
3543 const WorkItem &WI = *I;
3544 size_t LUIdx = WI.LUIdx;
3545 LSRUse &LU = Uses[LUIdx];
3546 int64_t Imm = WI.Imm;
3547 const SCEV *OrigReg = WI.OrigReg;
3549 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3550 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3551 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3553 // TODO: Use a more targeted data structure.
3554 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3555 const Formula &F = LU.Formulae[L];
3556 // Use the immediate in the scaled register.
3557 if (F.ScaledReg == OrigReg) {
3558 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
3559 // Don't create 50 + reg(-50).
3560 if (F.referencesReg(SE.getSCEV(
3561 ConstantInt::get(IntTy, -(uint64_t)Offset))))
3564 NewF.BaseOffset = Offset;
3565 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3568 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3570 // If the new scale is a constant in a register, and adding the constant
3571 // value to the immediate would produce a value closer to zero than the
3572 // immediate itself, then the formula isn't worthwhile.
3573 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3574 if (C->getValue()->isNegative() !=
3575 (NewF.BaseOffset < 0) &&
3576 (C->getValue()->getValue().abs() * APInt(BitWidth, F.Scale))
3577 .ule(abs64(NewF.BaseOffset)))
3581 (void)InsertFormula(LU, LUIdx, NewF);
3583 // Use the immediate in a base register.
3584 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3585 const SCEV *BaseReg = F.BaseRegs[N];
3586 if (BaseReg != OrigReg)
3589 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
3590 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
3591 LU.Kind, LU.AccessTy, NewF)) {
3592 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
3595 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
3597 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
3599 // If the new formula has a constant in a register, and adding the
3600 // constant value to the immediate would produce a value closer to
3601 // zero than the immediate itself, then the formula isn't worthwhile.
3602 for (SmallVectorImpl<const SCEV *>::const_iterator
3603 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
3605 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
3606 if ((C->getValue()->getValue() + NewF.BaseOffset).abs().slt(
3607 abs64(NewF.BaseOffset)) &&
3608 (C->getValue()->getValue() +
3609 NewF.BaseOffset).countTrailingZeros() >=
3610 CountTrailingZeros_64(NewF.BaseOffset))
3614 (void)InsertFormula(LU, LUIdx, NewF);
3623 /// GenerateAllReuseFormulae - Generate formulae for each use.
3625 LSRInstance::GenerateAllReuseFormulae() {
3626 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
3627 // queries are more precise.
3628 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3629 LSRUse &LU = Uses[LUIdx];
3630 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3631 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
3632 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3633 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
3635 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3636 LSRUse &LU = Uses[LUIdx];
3637 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3638 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
3639 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3640 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
3641 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3642 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
3643 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3644 GenerateScales(LU, LUIdx, LU.Formulae[i]);
3646 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3647 LSRUse &LU = Uses[LUIdx];
3648 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3649 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
3652 GenerateCrossUseConstantOffsets();
3654 DEBUG(dbgs() << "\n"
3655 "After generating reuse formulae:\n";
3656 print_uses(dbgs()));
3659 /// If there are multiple formulae with the same set of registers used
3660 /// by other uses, pick the best one and delete the others.
3661 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
3662 DenseSet<const SCEV *> VisitedRegs;
3663 SmallPtrSet<const SCEV *, 16> Regs;
3664 SmallPtrSet<const SCEV *, 16> LoserRegs;
3666 bool ChangedFormulae = false;
3669 // Collect the best formula for each unique set of shared registers. This
3670 // is reset for each use.
3671 typedef DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>
3673 BestFormulaeTy BestFormulae;
3675 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3676 LSRUse &LU = Uses[LUIdx];
3677 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
3680 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
3681 FIdx != NumForms; ++FIdx) {
3682 Formula &F = LU.Formulae[FIdx];
3684 // Some formulas are instant losers. For example, they may depend on
3685 // nonexistent AddRecs from other loops. These need to be filtered
3686 // immediately, otherwise heuristics could choose them over others leading
3687 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
3688 // avoids the need to recompute this information across formulae using the
3689 // same bad AddRec. Passing LoserRegs is also essential unless we remove
3690 // the corresponding bad register from the Regs set.
3693 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT,
3695 if (CostF.isLoser()) {
3696 // During initial formula generation, undesirable formulae are generated
3697 // by uses within other loops that have some non-trivial address mode or
3698 // use the postinc form of the IV. LSR needs to provide these formulae
3699 // as the basis of rediscovering the desired formula that uses an AddRec
3700 // corresponding to the existing phi. Once all formulae have been
3701 // generated, these initial losers may be pruned.
3702 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
3706 SmallVector<const SCEV *, 4> Key;
3707 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
3708 JE = F.BaseRegs.end(); J != JE; ++J) {
3709 const SCEV *Reg = *J;
3710 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
3714 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
3715 Key.push_back(F.ScaledReg);
3716 // Unstable sort by host order ok, because this is only used for
3718 std::sort(Key.begin(), Key.end());
3720 std::pair<BestFormulaeTy::const_iterator, bool> P =
3721 BestFormulae.insert(std::make_pair(Key, FIdx));
3725 Formula &Best = LU.Formulae[P.first->second];
3729 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
3730 if (CostF < CostBest)
3732 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
3734 " in favor of formula "; Best.print(dbgs());
3738 ChangedFormulae = true;
3740 LU.DeleteFormula(F);
3746 // Now that we've filtered out some formulae, recompute the Regs set.
3748 LU.RecomputeRegs(LUIdx, RegUses);
3750 // Reset this to prepare for the next use.
3751 BestFormulae.clear();
3754 DEBUG(if (ChangedFormulae) {
3756 "After filtering out undesirable candidates:\n";
3761 // This is a rough guess that seems to work fairly well.
3762 static const size_t ComplexityLimit = UINT16_MAX;
3764 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
3765 /// solutions the solver might have to consider. It almost never considers
3766 /// this many solutions because it prune the search space, but the pruning
3767 /// isn't always sufficient.
3768 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
3770 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3771 E = Uses.end(); I != E; ++I) {
3772 size_t FSize = I->Formulae.size();
3773 if (FSize >= ComplexityLimit) {
3774 Power = ComplexityLimit;
3778 if (Power >= ComplexityLimit)
3784 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
3785 /// of the registers of another formula, it won't help reduce register
3786 /// pressure (though it may not necessarily hurt register pressure); remove
3787 /// it to simplify the system.
3788 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
3789 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3790 DEBUG(dbgs() << "The search space is too complex.\n");
3792 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
3793 "which use a superset of registers used by other "
3796 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3797 LSRUse &LU = Uses[LUIdx];
3799 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3800 Formula &F = LU.Formulae[i];
3801 // Look for a formula with a constant or GV in a register. If the use
3802 // also has a formula with that same value in an immediate field,
3803 // delete the one that uses a register.
3804 for (SmallVectorImpl<const SCEV *>::const_iterator
3805 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
3806 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
3808 NewF.BaseOffset += C->getValue()->getSExtValue();
3809 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3810 (I - F.BaseRegs.begin()));
3811 if (LU.HasFormulaWithSameRegs(NewF)) {
3812 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3813 LU.DeleteFormula(F);
3819 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
3820 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
3824 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3825 (I - F.BaseRegs.begin()));
3826 if (LU.HasFormulaWithSameRegs(NewF)) {
3827 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3829 LU.DeleteFormula(F);
3840 LU.RecomputeRegs(LUIdx, RegUses);
3843 DEBUG(dbgs() << "After pre-selection:\n";
3844 print_uses(dbgs()));
3848 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3849 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3851 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3852 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
3855 DEBUG(dbgs() << "The search space is too complex.\n"
3856 "Narrowing the search space by assuming that uses separated "
3857 "by a constant offset will use the same registers.\n");
3859 // This is especially useful for unrolled loops.
3861 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3862 LSRUse &LU = Uses[LUIdx];
3863 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3864 E = LU.Formulae.end(); I != E; ++I) {
3865 const Formula &F = *I;
3866 if (F.BaseOffset == 0 || F.Scale != 0)
3869 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
3873 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
3874 LU.Kind, LU.AccessTy))
3877 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n');
3879 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3881 // Update the relocs to reference the new use.
3882 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3883 E = Fixups.end(); I != E; ++I) {
3884 LSRFixup &Fixup = *I;
3885 if (Fixup.LUIdx == LUIdx) {
3886 Fixup.LUIdx = LUThatHas - &Uses.front();
3887 Fixup.Offset += F.BaseOffset;
3888 // Add the new offset to LUThatHas' offset list.
3889 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3890 LUThatHas->Offsets.push_back(Fixup.Offset);
3891 if (Fixup.Offset > LUThatHas->MaxOffset)
3892 LUThatHas->MaxOffset = Fixup.Offset;
3893 if (Fixup.Offset < LUThatHas->MinOffset)
3894 LUThatHas->MinOffset = Fixup.Offset;
3896 DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
3898 if (Fixup.LUIdx == NumUses-1)
3899 Fixup.LUIdx = LUIdx;
3902 // Delete formulae from the new use which are no longer legal.
3904 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3905 Formula &F = LUThatHas->Formulae[i];
3906 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
3907 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
3908 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3910 LUThatHas->DeleteFormula(F);
3918 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3920 // Delete the old use.
3921 DeleteUse(LU, LUIdx);
3928 DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
3931 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
3932 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
3933 /// we've done more filtering, as it may be able to find more formulae to
3935 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
3936 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3937 DEBUG(dbgs() << "The search space is too complex.\n");
3939 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
3940 "undesirable dedicated registers.\n");
3942 FilterOutUndesirableDedicatedRegisters();
3944 DEBUG(dbgs() << "After pre-selection:\n";
3945 print_uses(dbgs()));
3949 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
3950 /// to be profitable, and then in any use which has any reference to that
3951 /// register, delete all formulae which do not reference that register.
3952 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
3953 // With all other options exhausted, loop until the system is simple
3954 // enough to handle.
3955 SmallPtrSet<const SCEV *, 4> Taken;
3956 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3957 // Ok, we have too many of formulae on our hands to conveniently handle.
3958 // Use a rough heuristic to thin out the list.
3959 DEBUG(dbgs() << "The search space is too complex.\n");
3961 // Pick the register which is used by the most LSRUses, which is likely
3962 // to be a good reuse register candidate.
3963 const SCEV *Best = 0;
3964 unsigned BestNum = 0;
3965 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3967 const SCEV *Reg = *I;
3968 if (Taken.count(Reg))
3973 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3974 if (Count > BestNum) {
3981 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3982 << " will yield profitable reuse.\n");
3985 // In any use with formulae which references this register, delete formulae
3986 // which don't reference it.
3987 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3988 LSRUse &LU = Uses[LUIdx];
3989 if (!LU.Regs.count(Best)) continue;
3992 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3993 Formula &F = LU.Formulae[i];
3994 if (!F.referencesReg(Best)) {
3995 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3996 LU.DeleteFormula(F);
4000 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
4006 LU.RecomputeRegs(LUIdx, RegUses);
4009 DEBUG(dbgs() << "After pre-selection:\n";
4010 print_uses(dbgs()));
4014 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
4015 /// formulae to choose from, use some rough heuristics to prune down the number
4016 /// of formulae. This keeps the main solver from taking an extraordinary amount
4017 /// of time in some worst-case scenarios.
4018 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4019 NarrowSearchSpaceByDetectingSupersets();
4020 NarrowSearchSpaceByCollapsingUnrolledCode();
4021 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4022 NarrowSearchSpaceByPickingWinnerRegs();
4025 /// SolveRecurse - This is the recursive solver.
4026 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4028 SmallVectorImpl<const Formula *> &Workspace,
4029 const Cost &CurCost,
4030 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4031 DenseSet<const SCEV *> &VisitedRegs) const {
4034 // - use more aggressive filtering
4035 // - sort the formula so that the most profitable solutions are found first
4036 // - sort the uses too
4038 // - don't compute a cost, and then compare. compare while computing a cost
4040 // - track register sets with SmallBitVector
4042 const LSRUse &LU = Uses[Workspace.size()];
4044 // If this use references any register that's already a part of the
4045 // in-progress solution, consider it a requirement that a formula must
4046 // reference that register in order to be considered. This prunes out
4047 // unprofitable searching.
4048 SmallSetVector<const SCEV *, 4> ReqRegs;
4049 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
4050 E = CurRegs.end(); I != E; ++I)
4051 if (LU.Regs.count(*I))
4054 SmallPtrSet<const SCEV *, 16> NewRegs;
4056 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
4057 E = LU.Formulae.end(); I != E; ++I) {
4058 const Formula &F = *I;
4060 // Ignore formulae which do not use any of the required registers.
4061 bool SatisfiedReqReg = true;
4062 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
4063 JE = ReqRegs.end(); J != JE; ++J) {
4064 const SCEV *Reg = *J;
4065 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
4066 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
4068 SatisfiedReqReg = false;
4072 if (!SatisfiedReqReg) {
4073 // If none of the formulae satisfied the required registers, then we could
4074 // clear ReqRegs and try again. Currently, we simply give up in this case.
4078 // Evaluate the cost of the current formula. If it's already worse than
4079 // the current best, prune the search at that point.
4082 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
4083 if (NewCost < SolutionCost) {
4084 Workspace.push_back(&F);
4085 if (Workspace.size() != Uses.size()) {
4086 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4087 NewRegs, VisitedRegs);
4088 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4089 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4091 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4092 dbgs() << ".\n Regs:";
4093 for (SmallPtrSet<const SCEV *, 16>::const_iterator
4094 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
4095 dbgs() << ' ' << **I;
4098 SolutionCost = NewCost;
4099 Solution = Workspace;
4101 Workspace.pop_back();
4106 /// Solve - Choose one formula from each use. Return the results in the given
4107 /// Solution vector.
4108 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4109 SmallVector<const Formula *, 8> Workspace;
4111 SolutionCost.Loose();
4113 SmallPtrSet<const SCEV *, 16> CurRegs;
4114 DenseSet<const SCEV *> VisitedRegs;
4115 Workspace.reserve(Uses.size());
4117 // SolveRecurse does all the work.
4118 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4119 CurRegs, VisitedRegs);
4120 if (Solution.empty()) {
4121 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4125 // Ok, we've now made all our decisions.
4126 DEBUG(dbgs() << "\n"
4127 "The chosen solution requires "; SolutionCost.print(dbgs());
4129 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4131 Uses[i].print(dbgs());
4134 Solution[i]->print(dbgs());
4138 assert(Solution.size() == Uses.size() && "Malformed solution!");
4141 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
4142 /// the dominator tree far as we can go while still being dominated by the
4143 /// input positions. This helps canonicalize the insert position, which
4144 /// encourages sharing.
4145 BasicBlock::iterator
4146 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4147 const SmallVectorImpl<Instruction *> &Inputs)
4150 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4151 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4154 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4155 if (!Rung) return IP;
4156 Rung = Rung->getIDom();
4157 if (!Rung) return IP;
4158 IDom = Rung->getBlock();
4160 // Don't climb into a loop though.
4161 const Loop *IDomLoop = LI.getLoopFor(IDom);
4162 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4163 if (IDomDepth <= IPLoopDepth &&
4164 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4168 bool AllDominate = true;
4169 Instruction *BetterPos = 0;
4170 Instruction *Tentative = IDom->getTerminator();
4171 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
4172 E = Inputs.end(); I != E; ++I) {
4173 Instruction *Inst = *I;
4174 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4175 AllDominate = false;
4178 // Attempt to find an insert position in the middle of the block,
4179 // instead of at the end, so that it can be used for other expansions.
4180 if (IDom == Inst->getParent() &&
4181 (!BetterPos || !DT.dominates(Inst, BetterPos)))
4182 BetterPos = llvm::next(BasicBlock::iterator(Inst));
4195 /// AdjustInsertPositionForExpand - Determine an input position which will be
4196 /// dominated by the operands and which will dominate the result.
4197 BasicBlock::iterator
4198 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4201 SCEVExpander &Rewriter) const {
4202 // Collect some instructions which must be dominated by the
4203 // expanding replacement. These must be dominated by any operands that
4204 // will be required in the expansion.
4205 SmallVector<Instruction *, 4> Inputs;
4206 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4207 Inputs.push_back(I);
4208 if (LU.Kind == LSRUse::ICmpZero)
4209 if (Instruction *I =
4210 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4211 Inputs.push_back(I);
4212 if (LF.PostIncLoops.count(L)) {
4213 if (LF.isUseFullyOutsideLoop(L))
4214 Inputs.push_back(L->getLoopLatch()->getTerminator());
4216 Inputs.push_back(IVIncInsertPos);
4218 // The expansion must also be dominated by the increment positions of any
4219 // loops it for which it is using post-inc mode.
4220 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
4221 E = LF.PostIncLoops.end(); I != E; ++I) {
4222 const Loop *PIL = *I;
4223 if (PIL == L) continue;
4225 // Be dominated by the loop exit.
4226 SmallVector<BasicBlock *, 4> ExitingBlocks;
4227 PIL->getExitingBlocks(ExitingBlocks);
4228 if (!ExitingBlocks.empty()) {
4229 BasicBlock *BB = ExitingBlocks[0];
4230 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4231 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4232 Inputs.push_back(BB->getTerminator());
4236 assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP)
4237 && !isa<DbgInfoIntrinsic>(LowestIP) &&
4238 "Insertion point must be a normal instruction");
4240 // Then, climb up the immediate dominator tree as far as we can go while
4241 // still being dominated by the input positions.
4242 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4244 // Don't insert instructions before PHI nodes.
4245 while (isa<PHINode>(IP)) ++IP;
4247 // Ignore landingpad instructions.
4248 while (isa<LandingPadInst>(IP)) ++IP;
4250 // Ignore debug intrinsics.
4251 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4253 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4254 // IP consistent across expansions and allows the previously inserted
4255 // instructions to be reused by subsequent expansion.
4256 while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP;
4261 /// Expand - Emit instructions for the leading candidate expression for this
4262 /// LSRUse (this is called "expanding").
4263 Value *LSRInstance::Expand(const LSRFixup &LF,
4265 BasicBlock::iterator IP,
4266 SCEVExpander &Rewriter,
4267 SmallVectorImpl<WeakVH> &DeadInsts) const {
4268 const LSRUse &LU = Uses[LF.LUIdx];
4270 // Determine an input position which will be dominated by the operands and
4271 // which will dominate the result.
4272 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4274 // Inform the Rewriter if we have a post-increment use, so that it can
4275 // perform an advantageous expansion.
4276 Rewriter.setPostInc(LF.PostIncLoops);
4278 // This is the type that the user actually needs.
4279 Type *OpTy = LF.OperandValToReplace->getType();
4280 // This will be the type that we'll initially expand to.
4281 Type *Ty = F.getType();
4283 // No type known; just expand directly to the ultimate type.
4285 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4286 // Expand directly to the ultimate type if it's the right size.
4288 // This is the type to do integer arithmetic in.
4289 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4291 // Build up a list of operands to add together to form the full base.
4292 SmallVector<const SCEV *, 8> Ops;
4294 // Expand the BaseRegs portion.
4295 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
4296 E = F.BaseRegs.end(); I != E; ++I) {
4297 const SCEV *Reg = *I;
4298 assert(!Reg->isZero() && "Zero allocated in a base register!");
4300 // If we're expanding for a post-inc user, make the post-inc adjustment.
4301 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4302 Reg = TransformForPostIncUse(Denormalize, Reg,
4303 LF.UserInst, LF.OperandValToReplace,
4306 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
4309 // Expand the ScaledReg portion.
4310 Value *ICmpScaledV = 0;
4312 const SCEV *ScaledS = F.ScaledReg;
4314 // If we're expanding for a post-inc user, make the post-inc adjustment.
4315 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4316 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
4317 LF.UserInst, LF.OperandValToReplace,
4320 if (LU.Kind == LSRUse::ICmpZero) {
4321 // An interesting way of "folding" with an icmp is to use a negated
4322 // scale, which we'll implement by inserting it into the other operand
4324 assert(F.Scale == -1 &&
4325 "The only scale supported by ICmpZero uses is -1!");
4326 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
4328 // Otherwise just expand the scaled register and an explicit scale,
4329 // which is expected to be matched as part of the address.
4331 // Flush the operand list to suppress SCEVExpander hoisting address modes.
4332 if (!Ops.empty() && LU.Kind == LSRUse::Address) {
4333 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4335 Ops.push_back(SE.getUnknown(FullV));
4337 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
4338 ScaledS = SE.getMulExpr(ScaledS,
4339 SE.getConstant(ScaledS->getType(), F.Scale));
4340 Ops.push_back(ScaledS);
4344 // Expand the GV portion.
4346 // Flush the operand list to suppress SCEVExpander hoisting.
4348 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4350 Ops.push_back(SE.getUnknown(FullV));
4352 Ops.push_back(SE.getUnknown(F.BaseGV));
4355 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
4356 // unfolded offsets. LSR assumes they both live next to their uses.
4358 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4360 Ops.push_back(SE.getUnknown(FullV));
4363 // Expand the immediate portion.
4364 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
4366 if (LU.Kind == LSRUse::ICmpZero) {
4367 // The other interesting way of "folding" with an ICmpZero is to use a
4368 // negated immediate.
4370 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4372 Ops.push_back(SE.getUnknown(ICmpScaledV));
4373 ICmpScaledV = ConstantInt::get(IntTy, Offset);
4376 // Just add the immediate values. These again are expected to be matched
4377 // as part of the address.
4378 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
4382 // Expand the unfolded offset portion.
4383 int64_t UnfoldedOffset = F.UnfoldedOffset;
4384 if (UnfoldedOffset != 0) {
4385 // Just add the immediate values.
4386 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
4390 // Emit instructions summing all the operands.
4391 const SCEV *FullS = Ops.empty() ?
4392 SE.getConstant(IntTy, 0) :
4394 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
4396 // We're done expanding now, so reset the rewriter.
4397 Rewriter.clearPostInc();
4399 // An ICmpZero Formula represents an ICmp which we're handling as a
4400 // comparison against zero. Now that we've expanded an expression for that
4401 // form, update the ICmp's other operand.
4402 if (LU.Kind == LSRUse::ICmpZero) {
4403 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
4404 DeadInsts.push_back(CI->getOperand(1));
4405 assert(!F.BaseGV && "ICmp does not support folding a global value and "
4406 "a scale at the same time!");
4407 if (F.Scale == -1) {
4408 if (ICmpScaledV->getType() != OpTy) {
4410 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
4412 ICmpScaledV, OpTy, "tmp", CI);
4415 CI->setOperand(1, ICmpScaledV);
4417 assert(F.Scale == 0 &&
4418 "ICmp does not support folding a global value and "
4419 "a scale at the same time!");
4420 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
4422 if (C->getType() != OpTy)
4423 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4427 CI->setOperand(1, C);
4434 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
4435 /// of their operands effectively happens in their predecessor blocks, so the
4436 /// expression may need to be expanded in multiple places.
4437 void LSRInstance::RewriteForPHI(PHINode *PN,
4440 SCEVExpander &Rewriter,
4441 SmallVectorImpl<WeakVH> &DeadInsts,
4443 DenseMap<BasicBlock *, Value *> Inserted;
4444 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
4445 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
4446 BasicBlock *BB = PN->getIncomingBlock(i);
4448 // If this is a critical edge, split the edge so that we do not insert
4449 // the code on all predecessor/successor paths. We do this unless this
4450 // is the canonical backedge for this loop, which complicates post-inc
4452 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
4453 !isa<IndirectBrInst>(BB->getTerminator())) {
4454 BasicBlock *Parent = PN->getParent();
4455 Loop *PNLoop = LI.getLoopFor(Parent);
4456 if (!PNLoop || Parent != PNLoop->getHeader()) {
4457 // Split the critical edge.
4458 BasicBlock *NewBB = 0;
4459 if (!Parent->isLandingPad()) {
4460 NewBB = SplitCriticalEdge(BB, Parent, P,
4461 /*MergeIdenticalEdges=*/true,
4462 /*DontDeleteUselessPhis=*/true);
4464 SmallVector<BasicBlock*, 2> NewBBs;
4465 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs);
4468 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
4469 // phi predecessors are identical. The simple thing to do is skip
4470 // splitting in this case rather than complicate the API.
4472 // If PN is outside of the loop and BB is in the loop, we want to
4473 // move the block to be immediately before the PHI block, not
4474 // immediately after BB.
4475 if (L->contains(BB) && !L->contains(PN))
4476 NewBB->moveBefore(PN->getParent());
4478 // Splitting the edge can reduce the number of PHI entries we have.
4479 e = PN->getNumIncomingValues();
4481 i = PN->getBasicBlockIndex(BB);
4486 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
4487 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
4489 PN->setIncomingValue(i, Pair.first->second);
4491 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
4493 // If this is reuse-by-noop-cast, insert the noop cast.
4494 Type *OpTy = LF.OperandValToReplace->getType();
4495 if (FullV->getType() != OpTy)
4497 CastInst::Create(CastInst::getCastOpcode(FullV, false,
4499 FullV, LF.OperandValToReplace->getType(),
4500 "tmp", BB->getTerminator());
4502 PN->setIncomingValue(i, FullV);
4503 Pair.first->second = FullV;
4508 /// Rewrite - Emit instructions for the leading candidate expression for this
4509 /// LSRUse (this is called "expanding"), and update the UserInst to reference
4510 /// the newly expanded value.
4511 void LSRInstance::Rewrite(const LSRFixup &LF,
4513 SCEVExpander &Rewriter,
4514 SmallVectorImpl<WeakVH> &DeadInsts,
4516 // First, find an insertion point that dominates UserInst. For PHI nodes,
4517 // find the nearest block which dominates all the relevant uses.
4518 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
4519 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
4521 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
4523 // If this is reuse-by-noop-cast, insert the noop cast.
4524 Type *OpTy = LF.OperandValToReplace->getType();
4525 if (FullV->getType() != OpTy) {
4527 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
4528 FullV, OpTy, "tmp", LF.UserInst);
4532 // Update the user. ICmpZero is handled specially here (for now) because
4533 // Expand may have updated one of the operands of the icmp already, and
4534 // its new value may happen to be equal to LF.OperandValToReplace, in
4535 // which case doing replaceUsesOfWith leads to replacing both operands
4536 // with the same value. TODO: Reorganize this.
4537 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
4538 LF.UserInst->setOperand(0, FullV);
4540 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
4543 DeadInsts.push_back(LF.OperandValToReplace);
4546 /// ImplementSolution - Rewrite all the fixup locations with new values,
4547 /// following the chosen solution.
4549 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
4551 // Keep track of instructions we may have made dead, so that
4552 // we can remove them after we are done working.
4553 SmallVector<WeakVH, 16> DeadInsts;
4555 SCEVExpander Rewriter(SE, "lsr");
4557 Rewriter.setDebugType(DEBUG_TYPE);
4559 Rewriter.disableCanonicalMode();
4560 Rewriter.enableLSRMode();
4561 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
4563 // Mark phi nodes that terminate chains so the expander tries to reuse them.
4564 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4565 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4566 if (PHINode *PN = dyn_cast<PHINode>(ChainI->tailUserInst()))
4567 Rewriter.setChainedPhi(PN);
4570 // Expand the new value definitions and update the users.
4571 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4572 E = Fixups.end(); I != E; ++I) {
4573 const LSRFixup &Fixup = *I;
4575 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
4580 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4581 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4582 GenerateIVChain(*ChainI, Rewriter, DeadInsts);
4585 // Clean up after ourselves. This must be done before deleting any
4589 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
4592 LSRInstance::LSRInstance(Loop *L, Pass *P)
4593 : IU(P->getAnalysis<IVUsers>()), SE(P->getAnalysis<ScalarEvolution>()),
4594 DT(P->getAnalysis<DominatorTree>()), LI(P->getAnalysis<LoopInfo>()),
4595 TTI(P->getAnalysis<TargetTransformInfo>()), L(L), Changed(false),
4597 // If LoopSimplify form is not available, stay out of trouble.
4598 if (!L->isLoopSimplifyForm())
4601 // If there's no interesting work to be done, bail early.
4602 if (IU.empty()) return;
4604 // If there's too much analysis to be done, bail early. We won't be able to
4605 // model the problem anyway.
4606 unsigned NumUsers = 0;
4607 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
4608 if (++NumUsers > MaxIVUsers) {
4609 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << *L
4616 // All dominating loops must have preheaders, or SCEVExpander may not be able
4617 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
4619 // IVUsers analysis should only create users that are dominated by simple loop
4620 // headers. Since this loop should dominate all of its users, its user list
4621 // should be empty if this loop itself is not within a simple loop nest.
4622 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
4623 Rung; Rung = Rung->getIDom()) {
4624 BasicBlock *BB = Rung->getBlock();
4625 const Loop *DomLoop = LI.getLoopFor(BB);
4626 if (DomLoop && DomLoop->getHeader() == BB) {
4627 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
4632 DEBUG(dbgs() << "\nLSR on loop ";
4633 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
4636 // First, perform some low-level loop optimizations.
4638 OptimizeLoopTermCond();
4640 // If loop preparation eliminates all interesting IV users, bail.
4641 if (IU.empty()) return;
4643 // Skip nested loops until we can model them better with formulae.
4645 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
4649 // Start collecting data and preparing for the solver.
4651 CollectInterestingTypesAndFactors();
4652 CollectFixupsAndInitialFormulae();
4653 CollectLoopInvariantFixupsAndFormulae();
4655 assert(!Uses.empty() && "IVUsers reported at least one use");
4656 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
4657 print_uses(dbgs()));
4659 // Now use the reuse data to generate a bunch of interesting ways
4660 // to formulate the values needed for the uses.
4661 GenerateAllReuseFormulae();
4663 FilterOutUndesirableDedicatedRegisters();
4664 NarrowSearchSpaceUsingHeuristics();
4666 SmallVector<const Formula *, 8> Solution;
4669 // Release memory that is no longer needed.
4674 if (Solution.empty())
4678 // Formulae should be legal.
4679 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), E = Uses.end();
4681 const LSRUse &LU = *I;
4682 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4683 JE = LU.Formulae.end();
4685 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4686 *J) && "Illegal formula generated!");
4690 // Now that we've decided what we want, make it so.
4691 ImplementSolution(Solution, P);
4694 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
4695 if (Factors.empty() && Types.empty()) return;
4697 OS << "LSR has identified the following interesting factors and types: ";
4700 for (SmallSetVector<int64_t, 8>::const_iterator
4701 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
4702 if (!First) OS << ", ";
4707 for (SmallSetVector<Type *, 4>::const_iterator
4708 I = Types.begin(), E = Types.end(); I != E; ++I) {
4709 if (!First) OS << ", ";
4711 OS << '(' << **I << ')';
4716 void LSRInstance::print_fixups(raw_ostream &OS) const {
4717 OS << "LSR is examining the following fixup sites:\n";
4718 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4719 E = Fixups.end(); I != E; ++I) {
4726 void LSRInstance::print_uses(raw_ostream &OS) const {
4727 OS << "LSR is examining the following uses:\n";
4728 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4729 E = Uses.end(); I != E; ++I) {
4730 const LSRUse &LU = *I;
4734 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4735 JE = LU.Formulae.end(); J != JE; ++J) {
4743 void LSRInstance::print(raw_ostream &OS) const {
4744 print_factors_and_types(OS);
4749 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4750 void LSRInstance::dump() const {
4751 print(errs()); errs() << '\n';
4757 class LoopStrengthReduce : public LoopPass {
4759 static char ID; // Pass ID, replacement for typeid
4760 LoopStrengthReduce();
4763 bool runOnLoop(Loop *L, LPPassManager &LPM);
4764 void getAnalysisUsage(AnalysisUsage &AU) const;
4769 char LoopStrengthReduce::ID = 0;
4770 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
4771 "Loop Strength Reduction", false, false)
4772 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
4773 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
4774 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4775 INITIALIZE_PASS_DEPENDENCY(IVUsers)
4776 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
4777 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4778 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
4779 "Loop Strength Reduction", false, false)
4782 Pass *llvm::createLoopStrengthReducePass() {
4783 return new LoopStrengthReduce();
4786 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
4787 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
4790 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
4791 // We split critical edges, so we change the CFG. However, we do update
4792 // many analyses if they are around.
4793 AU.addPreservedID(LoopSimplifyID);
4795 AU.addRequired<LoopInfo>();
4796 AU.addPreserved<LoopInfo>();
4797 AU.addRequiredID(LoopSimplifyID);
4798 AU.addRequired<DominatorTree>();
4799 AU.addPreserved<DominatorTree>();
4800 AU.addRequired<ScalarEvolution>();
4801 AU.addPreserved<ScalarEvolution>();
4802 // Requiring LoopSimplify a second time here prevents IVUsers from running
4803 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
4804 AU.addRequiredID(LoopSimplifyID);
4805 AU.addRequired<IVUsers>();
4806 AU.addPreserved<IVUsers>();
4807 AU.addRequired<TargetTransformInfo>();
4810 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
4811 bool Changed = false;
4813 // Run the main LSR transformation.
4814 Changed |= LSRInstance(L, this).getChanged();
4816 // Remove any extra phis created by processing inner loops.
4817 Changed |= DeleteDeadPHIs(L->getHeader());
4818 if (EnablePhiElim && L->isLoopSimplifyForm()) {
4819 SmallVector<WeakVH, 16> DeadInsts;
4820 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr");
4822 Rewriter.setDebugType(DEBUG_TYPE);
4824 unsigned numFolded =
4825 Rewriter.replaceCongruentIVs(L, &getAnalysis<DominatorTree>(),
4827 &getAnalysis<TargetTransformInfo>());
4830 DeleteTriviallyDeadInstructions(DeadInsts);
4831 DeleteDeadPHIs(L->getHeader());